Anima on M1 Max: tested 6 characters for a front-view + hair-intakes LoRA seed

Yesterday’s hair-intakes transplant experiment produced a recipe (J/AC) that grafts Nayuki’s big hair intakes onto a generic blonde character, but the output is not directly usable as LoRA training data.
Outputs are biased toward 3/4 view and the from-above AC composition; full front view is missing. LoRA training data needs a balanced mix of front, 3/4, and profile angles.

Three technical routes for getting Nayuki-style intakes in a front view:

1. Inpaint split job (front-view base image + only repaint the bangs region with intakes)
2. ControlNet OpenPose to force a front pose via skeleton
3. Switch the reference character to one whose hair intakes survive in a front-view composition

1 and 2 are too heavy on local hardware (M1 Max) to produce LoRA-scale dataset volumes.
This post goes with option 3: search Anima’s training distribution for a character whose hair intakes come out cleanly in a front view, and lock in that character’s recipe as the base for collecting LoRA training data.

Candidate selection criteria

Axes for picking a character likely to give a front view + hair intakes on Anima.

• Has hair intakes in the first place (trained under the Danbooru hair intakes tag)
• High ratio of front-view official art (idol-series stage art and character-select standing arts should bias toward front)
• Enough samples in Anima’s training data (too obscure → identity resolution is too low)
• Design consistency (characters whose intakes appear/disappear across eras are unstable)
• Female character (the base is a generic “girl” LoRA, so male references add extra body-feature removal work)
• Exclude FGO / Blue Archive / Azur Lane (yesterday’s V/Toki results showed heavy identity removal, or rigid franchise-specific motifs)

Candidates and rejection reasons, all in one table.

Character	Series	Verdict	Reason
Mogami Shizuka	iM@S Million Live	✅ Selected	Medium intakes, consistent design, plenty of front-view stage art
Yumemi Riamu	iM@S Cinderella Girls	✅ Selected	Sharp intakes, recent design, lots of Cinderella 10th-anniversary training data
Ange Katrina	Nijisanji	✅ Selected	Clear sharp intakes, popular Nijisanji VTuber
Sakura Kinomoto	Cardcaptor Sakura	✅ Selected	Bangs + sidelocks structure, 30 years of fan-art accumulation, consistent design
Kibasaki Ren	SideM	❌ Excluded	Male character, body features add extra removal
Takatsuki Yayoi	iM@S 765AS	❌ Excluded	Early (2005) design has no intakes, added later, training data mixes old/new and gets unstable
Tenkubashi Tomoka	iM@S Million Live	❌ Excluded	Sharp in doujin but iffy in official design, inconsistent
Kamio Misuzu	AIR	❌ Excluded	Intake on one side only, no left-right symmetry
Astolfo	Fate/Apocrypha	❌ Excluded	Treated as male, FGO makes identity removal heavy
Nero Claudius	Fate/Extra	❌ Excluded	Yesterday’s V experiment confirmed heavy removal
Minase Nayuki / Mizuse Akiko / Misaka Kaori	Kanon	△ On hold	Verified in last experiment, but the focus this time is “recent + front-view-heavy”, so set aside

Breakdown of the 4 selected characters.

Character	Role
Mogami Shizuka	Idol-style, front-view representative, certified by Million-Live Ps
Yumemi Riamu	Sharp-intakes recent idol slot
Ange Katrina	Recent VTuber slot, top-class intake strength
Sakura Kinomoto	Classic slot, representative of training-data accumulation

Three idols + one classic. Intake strengths span from medium to sharp.

Experiment phase 1: minimal prompt to check each candidate’s baseline

Each candidate is generated with the same minimal prompt (same procedure as yesterday’s Nayuki test), checking the following.

• Whether Anima recognizes the character (4/4 images that look like the actual person)
• Whether hair intakes appear, and roughly what size
• Composition bias (where it lands among front / 3/4 / side)

masterpiece, best quality, score_7, safe,
1girl, solo, [character_tag], [series_tag],
upper body, looking at viewer, simple background

Mogami Shizuka (mogami shizuka, idolmaster)

4/4 for character identity. Composition is 3/4-ish but settles near front.
Intakes are mild (medium, faintly forward-pointing locks on both sides of the bangs).
Million-Live uniform (navy jacket + white blouse + blue ribbon) is consistent.

Yumemi Riamu (yumemi riamu, idolmaster cinderella girls)

4/4 for identity, and all 4 are facing-viewer full front.
Intakes are the pink portions on both sides of the bangs forming a near-forward-pointing scoop, medium-ish.
Cyan is an inner-color gradient on the hair tips and is a separate decorative axis from intake structure.
Plenty of outfit motifs (devil horns, heart pattern, choker, pink + cyan frills) → heavy removal load.

Ange Katrina (ange katrina, nijisanji)

4/4 for identity, 4/4 are nearly front-facing (with a slight head tilt, not perfectly straight-on).
Both sides of the bangs have a clearly forward-pointing scoop, the strongest intake intensity among the 4 candidates.
Outfit motifs (red jacket + black + gold trim + triangle pendant + triangle clip) have many removable elements.

Sakura Kinomoto (kinomoto sakura, cardcaptor sakura)

4/4 for identity, grade-school Sakura + Tomoeda Elementary uniform (black + white sailor) + white headband-style headpiece + two red pom-poms.
Composition is closer to 3/4 with the face angled noticeably to the side; not a clean front.
Intakes come from near the top of the head (above the temples, near the hairline), with two locks pointing forward — small because of the short hair, but the structure is clearly there. An ahoge tags along.
Across all 4 images, hairstyle, outfit, and age impression don’t drift; Sakura’s identity is firmly baked into Anima.

Phase 1 summary

Character	Recognition	Intakes	Composition	Outfit removal load
Mogami Shizuka	✅	△ medium	3/4-ish front	Light (Million-Live uniform)
Yumemi Riamu	✅	◯ medium (pink portion)	✅ Full front	Heavy (horns, hearts, gradient hair)
Ange Katrina	✅✅	✅✅ clear	Nearly front (light head tilt)	Medium (red jacket + gold trim)
Sakura Kinomoto	✅	△ small	Tilted (face sideways)	Medium (CCS uniform)

The top pick is Ange Katrina. Intake intensity, front ratio, and recognition accuracy are all top.
Second is Yumemi Riamu (best front ratio, but intakes are a step weaker than Ange).
Shizuka and Sakura lean medium-intake and are put on hold for the purposes of moving into phase 2.

The Asakura Yume fallback isn’t needed (Ange and Riamu should give enough intensity).

Experiment phase 2: transplant onto AC composition with a J-recipe style

Take the phase-1 leaders (intake intensity + front ratio): Ange Katrina and Yumemi Riamu. Combine AC composition (from above, slightly turned to the side, almost front view) with the J recipe (NL reference + negatives for identity removal + weight 1.5).
Base prompt is the generic blonde used yesterday in J/AC. Replace the like minase nayuki from kanon (game) portion of the NL with the selected character, and add the character’s unique colors, outfit, and accessories to the negatives.

AC-Ange (via like ange katrina from nijisanji)

Add ange katrina, red hair, red jacket, white collar, gold trim, gothic outfit, star pendant, triangle pendant, hair clip to the negative.

4/4 holds the AC composition, 4/4 produce clear forward-pointing intakes on both sides of the bangs.
Compared to yesterday’s J (Nayuki version), intake intensity is equal or stronger, and it holds across all 4 seeds.
Red hair and red jacket are removable via the negatives, leaving a blonde + white sleeveless generic character.
However, the triangle clip is still present in 3/4 (the hair clip negative isn’t pulling enough weight, more specific tags like triangle hair clip are needed).
Bangs aren’t a clean blunt cut, more of a hime-cut-ish structure (bangs + split sides). Ange’s own bangs are split-type, so that’s bleeding through.

AC-Riamu (via like yumemi riamu from idolmaster cinderella girls)

Add yumemi riamu, pink hair, cyan hair, multicolored hair, gradient hair, devil horns, demon horns, heart pattern, heart ornament, frilly outfit, gothic lolita, choker to the negative.

4/4 holds the AC composition; crown braid is OK.
Intakes are noticeably weaker than AC-Ange — only as much as sidelock lift.
Outfit succeeds in genericizing to a white T-shirt by removing horns, heart pattern, frills.
However, cyan inner color on the hair tips leaks in all 4 seeds. The negative cyan hair, multicolored hair, gradient hair is not pulling enough weight.
On the 4th, eye color also leaks toward purple (Riamu’s violet/pink eyes seeping in).

The curious part: with Riamu alone (phase 1’s minimal prompt), the forward-pointing bangs (intake-like structure) showed up clearly, but the moment you mix that into the transplant prompt, the intakes weaken.
Ahoge goes the other way — not specified in the base prompt but persists in all 4 seeds.
Hypothesis: ahoge is a simple independent decorative element that rides naturally on top of a generic character context, while intakes require a structural change to the entire bangs and clash with the base blunt bangs specification.
Or Anima’s “intake feature” is strongly entangled with the Kanon-cluster prior internally, and a like riamu reference can’t reach into that prior.

Phase 2 results summary

There’s a clear difference in outfit learning structure between the two characters.

Character	Learning structure
Ange	Learning is concentrated on the VTuber Live2D/3D default outfit (red jacket family). Outfit motifs are rigid, but hair color / eye color are removed normally.
Riamu	Wears multiple outfits in Deresta → Anima learns multi-pattern outfits → harder to predict which one comes out, but they’re removed surprisingly easily. Hair gradient and eye color are sticky.

Evaluation as a LoRA training-data base.

Axis	Ange	Riamu
Intake intensity (post-AC)	✅✅ Clear forward scoop in 4/4	△ Faint in 4/4
Hair color removal	✅ Red → blonde OK	✗ Cyan gradient on tips
Eye color removal	✅ Blue retained	△ Purple leak on some seeds
Outfit removal	△ Triangle clip remains, needs extra negative	✅ Horns + heart removed OK
Composition stability	✅✅ 4/4 holds AC	✅✅ 4/4 holds AC

Overall Ange wins. Intake intensity is on another level and hair/eye color come off cleanly, finishing as a generic character.
Riamu drops out as a training-data candidate the moment its intakes go weak. Going to push Ange hard from here.

Experiment phase 3: does Ange give a true front view?

Now that Ange is the phase-2 pick, test in isolation whether full front + intakes + braid can coexist — essential for LoRA training data.
Yesterday’s M/O/P showed front + half braid + intakes don’t easily co-occur, but the reference character switch to Ange might change behavior.
Switch the composition spec to front view, facing viewer, upper body, looking at viewer, and add profile, from side, three quarters view to the negative.
NL reference and identity removal are the same as AC-Ange (drop the yesterday-M looking up line, add triangle hair clip to the negative).

Specified front view, facing viewer, but all 4 images have the face off-axis from slightly to clearly tilted, not the dead-front view I want for LoRA training data.
Intakes rise on both sides in 4/4, but the composition is closer to AC rather than dead front.
On top of that, 1/4 (F-03) leaked Ange’s red hair through the negative. Even with red hair negated, certain seeds slip through and identity leaks.
The triangle clip also remains on some seeds — Ange’s identity may be baked in more strongly than expected.

In short, Front-Ange alone can’t directly produce dead-front + intakes LoRA training data.
It can be used as part of the dataset, but it needs to be combined with other angles (especially left-right symmetric tilts / mirrored compositions).

At this point I can’t confirm Ange as the final pick. Front-Ange has a strong composition bias (rightward tilt), and a left-right symmetric training set will need either mirrored augmentation or another reference character mixed in.

There’s also a bangs problem.
The base character is supposed to have blunt bangs, but Front-Ange outputs all 4 with bangs in a center-parted split structure. Ange’s own bangs (center-parted + falling on both sides) are coming through the like ange katrina NL reference. Explicit blunt bangs has weak effect.
The base character is way off target, and the bangs structure mismatch must be resolved to use this for LoRA training data.

As an alternative or supplement to Ange, also test Shizuka — who already has half-up — riding the AC composition to pull intakes harder, in parallel.

Condition AD: Shizuka base with AC composition + weight 1.8 + blonde + add braid

Shizuka already has half-up in phase 1’s minimal prompt, blunt-ish bangs, and the Million-Live uniform is light as outfit motifs go. Going from “half-up → half braid up” only needs adding braid and should ride a generic character more naturally than Ange.
Swap the trailing NL to like mogami shizuka from idolmaster million live and raise the weight from 1.5 → 1.8. Add Shizuka’s black hair / Million-Live uniform removal elements to the negative.

4/4 blonde-ification succeeded (Shizuka’s black hair removable via negative). Half-up + blue ribbon structure is retained; outfit converged to the Million-Live blue-uniform cluster (blue dress / jumper skirt).
Blunt bangs come out cleanly in #4, while #1–3 still lean hime-cut. Even so, the trajectory is closer to target than the Ange reference (center-parted + crossed bangs).
Intakes remain weak to medium even at weight 1.8 — Shizuka’s own intakes are conservative in size, so they don’t grow as big as Ange.

Ange vs Shizuka tradeoff.

Axis	Ange base	Shizuka base
Intake intensity	✅ Strong	△ Medium
Bangs off-target	✗ Center-parted + crossed, off-target	✅ Tends to come out blunt-leaning
Outfit removal load	△ Triangle clip etc. remain	✅ Light, converges to blue uniform
Composition bias	✗ Right-tilt, can’t get dead front	△ 3/4-ish, with blunt seeds nudging front
Hair color removal	✅ Red → blonde	✅ Black → blonde

Overall Shizuka base is closer to the generic character I wanted to draw. Ange has top intake intensity but the other elements are too off-target to use as a training-data base.
Next: push Shizuka harder to see if intake size can grow.

Condition AE: push Shizuka base harder (weight 2.0 + NL reinforcement)

Keep AD’s configuration, raise (hair intakes:1.8) to (hair intakes:2.0), and reinforce the NL with “very prominent sharp forward-facing intake locks … clearly defined at the temples and pointing forward”.
Also add (blunt bangs:1.2) to push bangs further toward blunt, and add center parted bangs, parted bangs to the negative to suppress crossed bangs.

Intake size stays roughly flat vs AD; weight 2.0 brings no visible increase.
Bangs at (blunt bangs:1.2) + negative center parted bangs still settle hime-cut-leaning in 4/4. Can’t claim clear blunt-ification over AD. Half braid + blue ribbon comes out in 4/4.
Composition is 3/4-ish, not full front.

Shizuka’s intakes are pinned at “medium” in Anima’s training data, with a hard ceiling no matter how much you crank the weight.
Same shape as yesterday’s J2 result (weight 2.0 also hit a J-condition ceiling for intakes) — Shizuka version reproduced identically. To push intake strength to Ange levels, Shizuka reference alone has a limit.

Tradeoff that 1-character references can’t escape

So far AC composition + J-recipe-style approach is stuck on the following tradeoff depending on reference character.

Direction	Profile
Ange direction	Strong intakes + off-target bangs / outfit / composition
Shizuka direction	Intake ceiling = medium + bangs / outfit / composition okay

Pushing a single-character reference further won’t break this, so next decision is between hybrid or fix the front-forcing side internally on Anima.

Four candidates.

1. Hybrid: Keep Shizuka base structure, weakly mix in only the intake reference from Ange (write (like ange katrina:0.5) into the NL alongside, strongly negate Ange identity)
2. Ange + bangs override: Stay on Ange direction with (blunt bangs:1.3) + reinforced negative center parted bangs, parted bangs to suppress the off-target bangs
3. Compromise on Shizuka: Accept medium intakes for LoRA training data
4. Reinforce the front-view prompt stack: Instead of a single front view, stack multiple front-forcing tags (symmetrical face / facing forward / centered composition / straight-on angle, etc.) to physically override Anima’s composition bias

Condition AF: hybrid attempt (Shizuka base + Ange intake shape reference)

Rewrite the NL to a 2-character reference: “Her front hair is styled like mogami shizuka from idolmaster million live with blunt bangs and half-up hair, while the intake locks at the temples are shaped like ange katrina from nijisanji with sharp forward-facing scoops.” Add identity elements for both characters to the negative.

Bangs didn’t collapse as much as I expected; image quality feels marginally better.
But intake size shows no visible increase vs AE, and bangs are still hime-cut-leaning at AD/AE-comparable rates.
Look closer and a faint trace of crossed bangs remains. Ange’s center-parted bangs leak through as a residue from the like ange katrina reference — a piece of Ange’s identity seeps in. Doesn’t blow up the bangs entirely, but the trace is observable.
Hybrid isn’t the decisive answer; need to push one base (Ange or Shizuka) on a different axis.

Condition AG: reinforce the front-view prompt stack

Override the front-facing direction from the prompt side, hard. Instead of a single front view, explicitly tag symmetry, gaze, and posture separately to lock Anima’s judgment onto dead front.

Stack the following into the positive composition part.

front view, facing forward, looking straight ahead, symmetrical face, eye contact,
body facing forward, head facing forward, fully frontal pose,
centered composition, character in the middle, straight-on angle, upper body

Reinforce avoidance elements in the negative.

side view, profile, from side, turned head, three quarters view, three-quarter view, looking up, looking away

Base character is the same Shizuka reference as AE (like mogami shizuka from idolmaster million live + (hair intakes:2.0) + (blunt bangs:1.3)).

4/4 lined up: dead front + symmetric half braid + blue ribbon + intakes on both sides.
The 3/4-bias problem dropped in one shot. Image quality is directly usable as the “dead front” angle for LoRA training data.
Intake size stays medium, but composition and hairstyle details all hit target.
Bangs classify as blunt, but the strand texture is thin enough that I can’t call them clearly blunt — they sit on a borderline edge.
Pushing weight higher tips that border over to the hime-cut side. To make blunt clearer, add the strand texture on a different axis (thick bangs / locks of bangs tags, or describe the bundled texture in NL).

The hime-cut tip is interesting: Shizuka herself doesn’t have a hime cut. The actual Shizuka has blunt-leaning straight bangs + sidelocks on both sides — doesn’t meet the hime-cut definition.
Yet pushing like mogami shizuka NL hard surfaces hime-cut elements, which suggests Anima’s internal mogami shizuka training images include some hime-cut-style renderings that surface as the “last identity residue”.
Same shape as yesterday’s J2, where (hair intakes:2.0) shifted toward Nayuki-likeness — there the surfacing feature (intakes) was wanted, here the surfacing feature (hime cut) is not.
What surfaces when you push a base-character × reference-character combo hard is not easy to predict, and has to be confirmed empirically each time.

Only image 1 lost the outfit and trended toward chest exposure.
Adding nude, topless, bare shoulders, exposed breasts to the negative would suppress it — an accident in the manageable range.

The front-view approach finding: for Anima, “multiple tags explicitly tagging symmetry / gaze / posture” is more effective than “a single front view tag”.
Yesterday’s M condition with just front view, facing viewer had drifted tilted; stacking multiple tags flipped that to 4/4 dead front.

Even more important: half braid hair is visible in a dead-front view too, confirmed by AG.
Yesterday’s article framed “Anima thinks half-braid braids aren’t visible in front view” as a constraint and worked around it with W(22.5°) or Z(from above). Today’s AG showed that constraint can be broken with a stronger tag stack.
It’s not that Anima “can’t show a half braid in front view”, but that “with only a single front view tag, it drifts diagonally and tucks the braid to the back”. Lock down symmetry, gaze, and posture with separate tags and front + half-braid coexist. Yesterday’s hypothesis had a lower ceiling than I thought; there was more room on the prompt side to constrain front view strongly.

Condition AH: AG base + bangs / sidelock adjustment + outfit-dropout suppression

Compare AG outputs against the target reference image to address remaining gaps.
Target: “blunt bangs with clear strand texture” + “long sidelocks reaching down to the chest on both sides, framing the face (informally ‘antennae’, Danbooru-tag-wise close to a short-to-medium independent variant of sidelocks)” + “white shirt outfit”.

Changes from AG.

• (hair intakes:2.0) → (hair intakes:1.3) (reduce scoop, push side locks toward sidelock-like)
• Add (sidelocks:1.4) to reinforce independent locks on both sides
• Add forelock to the tag list
• Add “thick blunt bangs with visible strand separation, short sidelocks framing the face on both sides with distinct strand separation” to NL to spell out texture and antennae
• Rewrite Shizuka-intake description to “pointing downward rather than scooping forward, blending with the sidelocks”, tilting away from scoop toward downward independent locks
• Add nude, topless, bare shoulders, exposed breasts, no clothing to the negative to suppress outfit-dropout seeds
• Add white shirt to the positive

4/4 stabilized on dead front + symmetric + white T-shirt-like outfit. Outfit-dropout seeds: zero this round (negative reinforcement worked).
But the antennae (independent locks on both sides) don’t grow. Target has them reaching the chest; AH stops at jaw / neck.
Going dead-front also shortens hair overall in the rendering (3/4 view would have stretched antennae down below the face), which contributes.

Anima reads the output as “blonde Shizuka with a braid added”

A hypothesis emerges.
The moment you put like mogami shizuka NL into the base, Anima processes the output as “Shizuka derivative”, leaving Shizuka’s own feature mass (short antennae, blunt-leaning bangs, Million-Live outfit cluster) dominantly intact.
Bangs not coming out fully blunt, antennae not growing, outfit ending up white-T-shirt-ish (white-leaning simple after Million-Live uniform removal) — all of these are carried over from the Shizuka side.

The current approach reads to Anima as “Shizuka - black hair - Million-Live uniform + blonde + braid” as a vector op, and never becomes a pure “generic character with intakes”.
To grow antennae further, switch the NL reference to a long-sidelock character or push sidelocks weight harder. But pushing weight too far surfaces other elements (hime-cut, etc.) at a ceiling (confirmed in AE).

This is the limit of the character-reference-via-NL approach overall. To make Anima recognize “someone who isn’t Shizuka”, I need to advance to actually collecting LoRA training data and learning a separate character. Once I bake the new character into a LoRA using the AG-level prompts as a seed, the “Shizuka derivative” recognition on Anima’s side gets overwritten.

Condition AI: drop Shizuka NL reference, describe long sidelocks directly in NL

Drop the Shizuka NL reference entirely to clear AH’s “short antennae + Shizuka dependence”, and describe “long sidelocks reaching the chest” directly in NL.

NL:

Her front hair has thick blunt bangs with clear strand separation.
Long sidelocks frame the face on both sides, reaching down to her chest,
with distinct strand definition like floating antenna locks.
The intake locks at the temples blend with these long sidelocks pointing downward.

Strengthen (sidelocks:1.6), add (hime cut:1.3) to the negative, add short sidelocks short hair to the negative.

Blunt bangs + long antennae reach the target. Clear progress over AH.
But a new problem: ahoge stretches unusually long in all 4 seeds. 2–3 strands sticking up from the top of the head.
On top of that, apparent age shifted adult-leaning vs AH. Dropping Shizuka dependence may have shifted the base distribution to an “adult blonde long” cluster.
Prime suspect: the “floating antenna locks” string in NL. Anima may be reading antenna = ahoge.

Condition AJ: suppress ahoge + push toward teen

Two fixes from AI.

• Add teenager, high school student to the positive, change NL opening to “A teenage girl in her late teens”
• Add (ahoge:1.3), antenna hair, hair antenna, sticking up hair, mature, adult, old to the negative

teenager pulled apparent age down to teen side. AJ-2 is pretty close to the target reference (blue ribbon + white shirt + blunt + long antennae + half braid + blue ribbon).
But ahoge still persists in all 4 seeds. The “floating antenna locks” in NL likely outweighs the antenna hair negative.

Condition AK: full ahoge removal closes in on the target

From AJ-2’s structure, judged “the only fix left is ahoge”; tweak NL and negative.

• Delete “like floating antenna locks” from NL, rewrite to “hanging straight beside the face with clear strand separation”
• Beef the negative to (ahoge:1.5), antenna hair, antenna locks, hair antenna, sticking up hair, cowlick, stray hair, single strand sticking up
• Also add crossed bangs (suppress cross-trace on bangs)

Ahoge suppressed in 3/4, with a slight short ahoge remnant in 1/4.
AK-1 and AK-3 are within striking distance of the target reference image.

• ✅ Dead front
• ✅ Blunt bangs (some texture)
• ✅ Antennae (long sidelocks) reach the chest
• ✅ Half braid + blue ribbon
• ✅ White shirt outfit (close to target)
• ✅ Blonde
• ✅ No ahoge (in 3/4)
• △ Apparent age still slightly elevated
• △ Single cross strand on bangs creeps in by seed

Remaining gaps: “age slightly lower” and “more crossed-bangs suppression” — to be addressed by AL.

Condition AL: lower age + reinforce crossed-bangs suppression

Reinforce safe via (safe:1.3) in the positive (in Anima, safe in the positive seems to work better than NSFW in the negative), add young face to reinforce age. Add (crossed bangs:1.4), (center parted bangs:1.2), (mature:1.3), (adult:1.3), (old:1.2), milf with weights to the negative.

Teen feel came down, but intakes flattened and got absorbed by sidelocks. The balance of (sidelocks:1.6) reinforcement and (hair intakes:1.3) lets sidelocks take the intake territory. Outcome: “blonde girl with long sideburns”, losing the intakes I wanted.

Intakes and long sidelocks are supposed to be separate axes, but Anima’s sidelocks cluster seems to be subsuming intakes.

Condition AM: separate intakes and sidelocks on different axes in NL

Rewrite the NL to make intakes and sidelocks explicitly independent.

The intake locks at the temples are separate from the long sidelocks,
forming clear forward-pointing scoops above the sidelocks at each temple.

Also return weight to (hair intakes:1.6).

Blunt bangs + long sidelocks + half braid + teen face + blue-ribbon outfit are stable, but intakes are still absorbed by sidelocks and don’t rise. NL separation alone isn’t enough.

Condition AN: reinforce intakes + lower sidelocks + reorder NL

Raise intake weight to (hair intakes:1.8) and drop sidelocks to (sidelocks:1.4) to prioritize intakes. Also reorder NL so the intake description comes before sidelocks.

The output is “modern, sure, but not the spec I asked for”. Intakes do show but small, not a scoop type. Converging on a subtle idol-style intake. Target is a “clear forward-pointing scoop” type, so switch to mixing another reference character to inject scoop traits.

Condition AO: Shizuka + Chitanda Eru hybrid

Chitanda Eru is the heroine of Kyoto Animation’s Hyouka, a high-school girl with black long hair + blunt bangs + long side hair framing the face. She doesn’t have intakes proper, but her side-lock bundles reach the chest, which is close to the “long antennae” I want.

Idea origin: Eru has “blunt bangs + long sides (but the side locks are independent of intakes, just falling from the bangs)”, and Shizuka has “intake-origin but a bit short”. Combining the two should give an “intake origin + grows long” antennae of the right shape — user hypothesis.

NL with both characters referenced:

Her front bangs and side hair are styled like chitanda eru from hyouka,
with thick blunt bangs and long sidelocks reaching down to her chest.
The intake locks at the temples are styled like mogami shizuka from idolmaster million live,
originating between the bangs and sidelocks and extending downward together with the sidelocks.

Add chitanda eru, mogami shizuka, sailor uniform, sailor collar, idolmaster uniform to the negative.

Blunt bangs + long sidelocks achieved, but Eru’s bangs-to-sidelocks one-piece structure dominates and the intake-origin separation weakens. Eru’s structure “flows directly down from the bangs”, weakening Shizuka’s “rises from intake origin”.

Condition AP: Nayuki + Chitanda Eru hybrid

Switch to Nayuki for the intake reference instead of Shizuka. Nayuki is intake-strong, so combining with Eru’s long sidelocks should produce “long-from-intake”.

Closer to spec than AO, but the Eru NL reference pulls the overall art style toward Kyoto-Animation. The “Eru feel” is too strong. Eru’s influence is larger than expected.

Crossed bangs vanished in 1/4 and stayed in 3/4 — at least there’s a change.

Condition AQ: Shizuka + Nayuki hybrid (Eru excluded)

Drop Eru’s Kyoto-Animation vibe and stack Shizuka base (bangs / half-up) + Nayuki (strong intake scoop).

Now Shizuka came out too strong and shifted toward hime cut. Pushing the Shizuka NL hard again resurfaced the “Shizuka hime-cut residue” first seen in AE. Antennae also shortened; pulled toward Shizuka’s features.

Condition AR: Nayuki-only reference + long sidelocks expressed via weight

Pivot. Drop Eru and Shizuka both, narrow NL reference to Nayuki only. Express long sidelocks through (long sidelocks:1.5) weight and NL description alone.

NL:

The intake locks at the temples form prominent forward-pointing scoops
like minase nayuki from kanon (game),
extending downward into long sidelocks reaching her chest, hanging straight beside the face.

Things balanced out here. 4/4 lined up the following.

• ✅ Dead front composition
• ✅ Blunt bangs
• ✅ Antennae (long sidelocks) reaching shoulder-to-chest
• ✅ Half braid + blue ribbon on both sides
• ✅ White shirt outfit (close to target)
• ✅ Blonde
• ✅ Teen face
• ◯ Medium intake rising above the sidelocks

After mixing all kinds of hybrids, narrowing NL reference to Nayuki only ended up the cleanest result. Using Eru / Shizuka NL alongside each other has each character’s identity seep through and pulls the art style toward the reference character. With Nayuki alone, only the intake feature of the “Kanon-cluster prior” rides cleanly.

But crossed bangs still remain in 3/4 of 4 seeds.

Condition AS: narrow down the cause of crossed bangs (add hair between eyes to negative)

Narrow down the cause. Since the crossed bangs negative isn’t producing effect, hypothesize that the Danbooru standard tag hair between eyes (the single central front strand) is the real name. Put (hair between eyes:1.4) strongly in the negative. Also drop bangs weight to (blunt bangs:1.1).

AS still leaves crossed bangs in 3/4. The hair between eyes negative + (blunt bangs:1.1) relaxation combo doesn’t reach the cause either.

Condition AT: drop forelock tag + intake 1.5 + NL simplification

Another suspect: forelock. May be summoning the forward strand. Drop forelock, simplify NL to a minimum, drop intake weight to 1.5.

AR/AS’s 3/4 cross drops to AT’s 2/4 cross — a small improvement. forelock removal + intake reduction has a small cross-suppression effect, but doesn’t eliminate them.

Condition AU: reinforce bangs-straight in NL + change lighting

Crosses still remain; another angle. Spell out the straight quality of bangs in NL, and specify “uniform lighting across” to reduce dimensional contrast. User pointed out “I can’t tell if that’s hair or a streak of light” — keeping open the possibility that lighting adjustment removes them.

Add soft lighting, studio lighting, even lighting, front lighting to the positive, and harsh shadows, dramatic lighting, side lighting, strong shadows on bangs to the negative. Add “Her bangs are completely straight, flowing parallel down from the hairline like neatly cut blunt bangs, with no hair strands crossing or overlapping each other across the forehead.” to NL.

AU is worse than AT (3/4 cross). Lighting change + bangs-straight NL reinforcement isn’t the deciding factor.

What’s been confirmed so far: a single crossed-bangs strand can’t be fully erased by prompt tuning. Candidate causes — none confirmed:

• Co-occurring decorative byproduct of the “blonde + blunt bangs + intakes + half-up” cluster in Anima’s base distribution
• Training data contains a high density of images with “one strand flowing from the parting”
• Some seeds render it as a highlight / light streak

Inpainting only the bangs region as post-processing is a route, but on local hardware (M1 Max) it’s too heavy to scale to LoRA dataset volumes.
The practical path is visually discarding crossed-bangs seeds at generation time — collecting training data at a 1–2-out-of-4 acceptance rate.

Conditions AV–AX: trial-and-error on lighting to drop face shadow

AR captured composition and features, but shadows fall on cheeks / chin / neck. For LoRA training material, it’s likely to learn “back-lighting bias” before “character features”. Switch to lighting instructions to suppress it.

AV adds front lit, key light from front, no backlight to positive and backlit, rim lighting to negative.

AW adds ambient light, soft diffuse light, soft front lighting.

AX adds high-key lighting to the positive.

AV/AW/AX all leave shadows on cheeks / chin while only the chest and top of head are bright. Even with front lit and high-key, the light source stays back-lit. Likely an Illustrious-base bias choosing “pretty-girl face = backlight / side light”.

Condition AY: drop all lighting instructions, keep only white background

Hypothesis flip: maybe the lighting instructions themselves are counterproductive. Pull every light-source tag and keep only (white background:1.4), simple background. Also remove lighting-related items from the negative.

AY-02 also dropped the crossed bangs and is the lightest face shadow in the AY series. Removing all lighting instructions yields lighter face shadows than inserting them.

Two issues remain.

• Top of head is right at the frame’s upper edge, with hair tips flowing out of frame
• Faint face shadow remains (Anima/Illustrious base model itself biases the rendering toward including shading)

Condition AZ: secure headroom + describe front face lighting in NL

Add (full head visible:1.3), (headroom above head:1.3), head and shoulders shot, space above hair to composition. To suppress face shadow, add flat shading, anime cel shading, soft diffuse light, (even lighting on face:1.3), (no shadows on face:1.3) to positive and (shadows on face:1.4), (face shadow:1.4), (cheek shadow:1.3) etc. to negative.

Headroom came out but face shadows don’t release as much as hoped. Showed the prompt to another AI (Codex), and the feedback landed.

• flat shading and anime cel shading contradict (the former = no shading, the latter = 2–3 distinct tones). SDXL family tends to default to “add some shadow” on contradictory prompts
• Triple-stacking upper body + head and shoulders shot + head to chest framing framing tokens pulls the camera into a pulled-back + backlit-ish composition
• (safe:1.3) on Illustrious also affects lighting and easily summons “holy backlight”. 1.0 or drop entirely is enough
• The quality tags themselves (masterpiece, best quality, score_7) push lighting presentation harder on Illustrious. For LoRA training material, drop or weaken quality tags

Condition BA: drop quality tags + drop all presentation + full-power shadow-killing negative

Switch the strategy to reflect all of Codex’s points. Drop masterpiece, best quality, score_7, (safe:1.3), the contradictory flat shading, and the triple-stacked head and shoulders shot, head to chest framing. Add flat color, no shadows, even lighting, ambient lighting, plain lighting, studio lighting, official art style, clean lineart to push the rendering toward flat. Add shadows, cast shadow, harsh shadow, dark face, backlight, rim light, contre-jour, dramatic lighting, chiaroscuro, strong contrast, vignette, bloom, lens flare, depth of field to the negative.

BA-04 is the final form this round. Cheek / chin shadows are clearly lighter than AY/AZ and the rendering goes flatter. Dead front, head-to-chest in frame, headroom above, no crossed bangs, blue ribbons on both sides, blunt bangs, long sidelocks, teen face — all elements line up.

As a side effect, the 4 BA images now show a blue ribbon at the collar. The blue ribbon token migrating to the collar instead of the hair is a thing, but for LoRA training material that’s just outfit variation, not harm.

BA configuration adopted as LoRA training material

The adopted configuration is BA.

Element	Content
NL reference	Started with like minase nayuki from kanon (game), but BA drops NL entirely and uses Danbooru tags only (intake summoned via (hair intakes:1.5) weight)
No quality tags	Drop masterpiece, best quality, score_7 family entirely (suppress the side effect of Illustrious-family pushing lighting presentation harder)
No safe tag	Drop (safe:1.3) (suppress the “holy backlight” side effect)
Composition	Just upper body, front view, looking at viewer. No triple-stacked framing tokens
Rendering	flat color, no shadows, even lighting, ambient lighting, plain lighting, studio lighting, official art style, clean lineart
Background	simple background, white background
Character traits	blonde hair, long hair, blue eyes, blunt bangs, (hair intakes:1.5), long sidelocks, half updo, braid, blue ribbon, white shirt
Negative reinforcement	shadows, cast shadow, harsh shadow, dark face, backlight, rim light, contre-jour, dramatic lighting, chiaroscuro, strong contrast, vignette, bloom, lens flare, depth of field — kill shading presentation with full force

Up until partway through, AR (Nayuki-only reference + front-view tag stack) was the leading pick, but face shadow falling on the cheek/chin meant the LoRA material would carry a shading bias — visible in AV/AW/AX. AY confirmed “dropping lighting instructions tilts flatter”. Codex’s feedback identified “the quality tags themselves push presentation”. BA dropped quality tags + added shadow-killing negative, and face shadow reduced significantly.

Hybrid routes (Shizuka + Eru, Nayuki + Eru, Shizuka + Nayuki) were tried partway, but Eru pulls Anima’s entire art style toward Kyoto-Animation and Shizuka surfaces the hime-cut residue. The takeaway: narrow NL to one character, or drop NL entirely and configure with Danbooru tags only, ends up most stable.

The remaining crossed-bangs problem

Across AR/AS/AT/AU/BA, the “crossed bangs” strand running diagonally across the forehead can’t be fully erased via prompt tuning. The cause is unconfirmed:

• Co-occurring decorative byproduct of the “blonde + blunt bangs + intakes + half-up” cluster in Anima’s base distribution
• Training data contains a high density of “one strand flowing from the parting”
• Some seeds render it as a highlight / light streak

Adding crossed bangs, hair between eyes, stray hairs, parted bangs, asymmetrical bangs with weights to the negative still leaves crosses in 2–3 out of 4. Switching to a workflow that visually rejects crossed-bangs seeds at generation time — collecting training material at a 1–2-out-of-4 acceptance rate. Once only crossed-bangs-free seeds are fed to LoRA training, the LoRA bakes in “no crossed bangs” and the problem resolves itself post-training.

References

• Yesterday’s hair-intakes transplant experiment
• Anima negative-prompt cleanup
• Rewriting LoRA training captions for WAI-Anima
• Character LoRA pushed to 12,000 steps
• Danbooru hair intakes wiki