Can MOSS-Image Keep a Side Ponytail Occluded in Right Profiles
Contents
In my previous MOSS-Image caption audit post, rebuilding all 70 training images and captions still left a stubborn failure: in right-facing profiles, the hair tie of the side ponytail that should be hidden behind Kana’s head kept appearing on the visible side.
Throughout this post, a “right profile” means her nose points toward the right edge of the image, so her right cheek and right ear face the viewer. Because Kana wears her side ponytail on her anatomical left, a right profile should place the hair tie on the far side of her head, with at most the top edge of the bound bundle peeking past the silhouette.
Using the exact same 70 images on a local EVO-X2, I tested whether changing how the captions describe the hairstyle and camera viewpoint would make the hair tie stay occluded in right profiles.
Test Setup and Baseline Context
Both training and inference ran on a local EVO-X2 PC (AMD ROCm). MOSS-Image provides a pretrained base model (Base, the foundation model) and a supervised fine-tuned variant (SFT, tuned to follow instructions more closely). Character LoRA adapters (small add-on weights for a character’s face and hairstyle) were trained on Base.
For evaluation renders, I used two setups: SFT paired with a 4-step distillation LoRA (an auxiliary weight that compresses generation into 4 denoising steps for fast previews—because Base takes 30–50 steps per image), and plain Base with no LoRAs loaded at all to isolate what the underlying Base model does on its own without SFT or the 4-step LoRA. In this post, “captions” refer to the text descriptions paired with training images, while “prompts” refer to the input text used during image generation. Here, I rewrote the training captions and also fed matching hairstyle phrasing into the inference prompts.
| Item | Configuration |
|---|---|
| Machine | Local EVO-X2 PC (AMD ROCm) |
| Training Model & Params | MOSS-Image Base, 1024px, LoRA rank 16 / alpha 16, batch size 1, AdamW, lr 0.0001, training seed 42 |
| Fixed Training Rules | Text encoder weights frozen; horizontal flipping and caption dropout (randomly dropping caption tokens during training) disabled |
| Inference (LoRA Comparisons) | MOSS-Image SFT + 4-step distillation LoRA (weight 1.0 each), 768×768, Euler sampler, 4 steps, CFG (prompt-adherence scale) 1.0 |
| Inference (Plain Base) | MOSS-Image Base only (no LoRAs), 768×768, Euler, 30 steps, CFG 3, flow shift (noise-schedule weighting) 5, empty negative prompt |
| Fixed Comparison Controls | White T-shirt, light grey background; random seeds 73251, 50173, 89014 left-to-right unless noted |
Our target character, Kana (trigger word kanachan), wears a side ponytail (tied hair bundle) secured with a blue scrunchie (hair tie) on her upper anatomical left, while the remaining lower hair falls straight down to the base of her neck as a short bob. Following v1–v3 in the previous post, take an earlier 500-step run trained with image-relative captions as v4 (with intermediate v5 omitted here); v4 placed the front-view hair tie on the right side, yet the loose back hair stretched down to chest length and right profiles still drew the hair tie on the near side.
Keeping the 70 PNG images unchanged, I trained two new 500-step LoRAs from scratch using the same Base initial weights: v6, which rewrites the hairstyle prose relative to her crown and left ear, and v7, which adds explicit viewpoint sentences and a tag list. Below are four training images from the dataset: front (kana_0001), left profile (kana_0024), rear (kana_0004), and right profile (kana_0034, the visual target reference used throughout later comparisons) from left to right.

Switching from Canvas Left/Right to Anatomical References
Captions up through v4 relied on canvas coordinates, writing A side ponytail on the image right for front views and A side ponytail on the image left for rear views. However, viewing the same head from front versus back flips canvas left and right. In the left profile (kana_0024), the scrunchie sits well above and slightly behind her left ear, and in rear views the remaining bob covers her nape beneath the knot.
To avoid canvas-coordinate flips, I rewrote the hairstyle description relative to her crown and left ear. The new hairstyle block for kana_0001 (front view) explains that only the upper hair on her left is gathered above and behind her left ear with a blue scrunchie, while the lower layer forms a short bob at her nape, the tail hangs beside her left cheek, and one curled ahoge (a stray curved hair strand sticking up from the crown) rises from the top:
Only the upper layer of hair on her left is swept back into a single side ponytail. The tie sits just below the crown on the upper left side of her head, well above and slightly behind her left ear. The lower hair falls freely over her nape in a short bob, with tapered ends at the base of her neck. A blue scrunchie circles the base. The tail arches over it and hangs beside her left cheek, separate from the loose hair framing her jaw. One curved lock sticks up from the crown.
Even with the root anchored, a tail that hangs beside her cheek in a standing front portrait (kana_0001) swings outward away from the bob in a running pose (kana_0037). Writing “hangs beside her left cheek” on a running frame would contradict the image, so only the tail placement phrasing was split for action poses.
Inspecting What Is Visible in Right-Profile Training Images
Before rewriting the captions, I selected 10 training images covering varied angles and character scales and sent them to Qwen4 (a vision-language model) without any existing captions to check how a visual model reads the hairstyle on its own. Qwen4 identified the high left-side placement and loose lower hair in left profiles and rear views, but flipped the facing direction on a small full-body uniform profile. I therefore zoomed into the head of each image and verified hair-tie visibility manually before editing the captions.
| Training Image | Visible Details in Source Image | Rule Applied to New English Caption |
|---|---|---|
Left profile (0024) | Blue fabric folds and a tail curving back, with a clear gap from the bob | Hair sweeping above her left ear to the tie, then curving down separately |
Right profiles (0017, 0034) | Only the upper tip of the tail peeks past the silhouette; no blue scrunchie visible | Hair tie is on the far side of her head, with only the top edge peeking at the silhouette |
Front-right three-quarter (0035) | Thin sliver of blue fabric past the crown with a small hair tuft above | Part of the hair tie is visible |
Full-body front (0018) | Small tail is recognizable, but fastener color and shape are hard to read | Position of the tail only; omitted specific fastener details |
Notably, out of all 70 training images, only 0017 (a full-body shot) and 0034 are strict right profiles where the hair tie is occluded. I carried over composition, clothing, expression, and background lines from the previous dataset, fixed incomplete sentence fragments in two files, and finalized the 70-pair v6 dataset.
Training v6 for 500 Steps with Detailed Prose
The 500-step v6 training loop from Base finished in 5 hours, 2 minutes, and 24 seconds. For evaluation renders, each angle used the matching hairstyle prose from its training caption, except for a “no-hairstyle front” check that kept the trigger kanachan, clothing, expression, and background while removing hairstyle sentences from the prompt. What mattered most in the right profile was not scrunchie color, but whether the bound bundle stayed hidden behind the silhouette.
| View | Prompted Hairstyle | Results Across 3 Seeds (73251, 50173, 89014) |
|---|---|---|
| Front | Upper-left tie hanging beside left cheek with short nape bob | All 3 placed the blue scrunchie on her upper left with a neck-length bob |
| Left profile | Hair swept above left ear with gap between tail and bob | All 3 showed the scrunchie above and behind the ear, separating tail and bob |
| Right profile | Tie hidden on far side, only top edge of bundle peeking out | All 3 faced right with a short nape, but 0/3 hid the tie; brown or dark fasteners appeared on the near side |
| Rear | Tie below crown on her left beside loose lower hair | All 3 placed the blue scrunchie on her upper left with the bob covering her nape |
| Full-body front | Same hairstyle prose as front close-up | All 3 placed a single ponytail on her upper left with a bob ending above the shoulders |
| No-hairstyle front | Character trigger kept, hairstyle sentences removed | All 3 reverted to a plain brown bob with no side ponytail, scrunchie, or ahoge |
Here are the six compositions in the same order as the table (seeds 73251, 50173, and 89014 from left to right). Front and left profile renders produced the scrunchie placement and separated bob as intended.


In the right profile, however, even without the blue scrunchie color, thin dark bands appeared on the center and right seeds, and all three drew the hair tie on the visible near side.

Rear and full-body front renders preserved her left-side placement even at smaller head scales, while removing hairstyle text from the prompt reduced the output to a plain brown bob despite keeping kanachan.



Removing the Character LoRA in the Right Profile
Since the right-profile prompt explicitly stated that the hair tie is hidden on the far side yet it still appeared in front, I kept the exact same prompt and seeds and unloaded only the v6 character LoRA (leaving SFT and the 4-step distillation LoRA active).
| Right-Profile Setup | Outcome |
|---|---|
| With 500-step character LoRA v6 (same images as above) | All 3 faced right, but drew the hair tie and long tail on the near side |
| Without character LoRA (SFT + 4-step LoRA only) | Hair color and facial style changed, yet all 3 still drew a thin band and long tail on the near side |
The top strip repeats the character-LoRA renders for reference, and the bottom strip shows the no-character-LoRA baseline. Even with the character LoRA removed, the same visible-tail failure persisted under this prompt.


Comparing Against the Earlier Canvas-Based v4 Run
Lining up source image kana_0034, the earlier 500-step v4 run, and v6 at seed 73251 from left to right shows that v4 lifted the back hair to expose the nape, whereas v6 kept the nape bob. Still, both generated a long tail hanging outside the silhouette that does not exist in 0034. Because the inference prompts also differed between v4 and v6, I stopped v6 at 500 steps and moved on to adding explicit tags and viewpoint sentences.

Training v7 with Explicit Viewpoint Sentences and Tags
Because detailed prose alone in v6 still exposed the tie in right profiles, I built v7 by keeping v6’s prose and adding left side ponytail to all 70 images, along with a sentence mapping anatomical left/right to canvas coordinates (front/rear) or specifying which cheek and ear face the viewer (profiles). For the two right profiles (0017, 0034) and one full-body shot (0018) where the blue scrunchie is obscured, I omitted the blue scrunchie tag.
| Training Image | Added Viewpoint Sentence | Excerpt from Tag List |
|---|---|---|
Front (0001) | Her face is seen from the front, so her own left is on the viewer’s right. | left side ponytail, short hair, ahoge, blue scrunchie, portrait, front view |
Rear (0004) | The viewer sees the back of her head and no face. Her own left is on the viewer’s left. | left side ponytail, short hair, ahoge, blue scrunchie, portrait, from behind |
Right profile (0034) | Her right cheek and right ear face the viewer in profile. The left side of her head is on the far side. | left side ponytail, short hair, ahoge, portrait, profile, from side |
Resuming from a v6 checkpoint would change the total training steps, so v7 was trained fresh for 500 steps from the same Base initial weights with training seed 42.
Six-View Results on v7
The v7 training run completed in 5 hours, 11 minutes, and 25 seconds. Inference prompts likewise included the matching viewpoint sentence and tags (adjusted for the light grey background, while the no-hairstyle check reused v6’s prompt).
| View | Inference Specification | Placement & Shape on v7 |
|---|---|---|
| Front | Face visible; her own left is on viewer’s right | 3/3 placed the blue scrunchie on the upper right |
| Left profile | Left cheek and ear in front; tie on near side | 3/3 showed the scrunchie above/behind the ear with a short nape bob |
| Right profile | Right cheek and ear in front; her left on far side | 3/3 tied the hair high at the back like a standard ponytail; 0/3 matched 0034 |
| Rear | Back of head visible; her left on viewer’s left | 3/3 placed the blue scrunchie on the upper left |
| Full-body front | Same left/right mapping plus full-body standing tags | 3/3 placed the bundle on the viewer’s right |
| No-hairstyle front | Both hairstyle prose and tag list removed | 3/3 produced a plain bob without ponytail or scrunchie |
Ordered left to right by seeds 73251, 50173, and 89014 as before, front and left profiles placed the blue scrunchie on her anatomical left. In right profiles, however, all three drew a high back-of-head ponytail—still keeping some loose hair below, yet forming a completely different silhouette from 0034. Checking only scrunchie color or left/right presence would miss that structural error.



Rear and full-body front renders again kept the side ponytail on her anatomical left, while stripping hairstyle text and tags left only the bob even with kanachan present.



Side-by-Side Right Profile Controls Between v6 and v7
To separate training differences from prompt differences, I ran both v6 and v7 across three prompt variants (shown as row labels in the comparison grid below) using seeds 73251, 50173, and 89014:
plain: v6’s original hairstyle prose onlyhybrid: the same prose plus the viewpoint sentence and tag listgeneric(v7 only):hybridwithleft side ponytailreplaced byside ponytailin the tag list while keeping the her-left prose intact
Every row faced right, and while fasteners shifted between hair-colored bands and black elastics, every single row and seed drew a long tail on the near side instead of matching 0034.

Describing Only What Is Visible from the Right Without Extra Training
Since adding tags did not hide the tail, I replaced only the hairstyle paragraph at inference time without retraining. Keeping the right-profile framing, white T-shirt, light grey background, and three seeds fixed (with no tags), I switched to three sentences describing strictly what is visible in 0034: a curled ahoge over the crown, a small fold of hair peeking just behind the crown on the far outline, and a continuous curtain of hair behind the visible ear down to the nape:
A single curled lock rises over the crown. Just behind the crown, a small fold of hair peeks above the far outline of her head. The hair behind her visible ear falls in a loose, continuous curtain over the back of her neck, ending in tapered points at the base of her neck.
For v7, I also tested adding a fourth sentence stating that the upper strands above her visible right ear sweep diagonally back and up toward that small fold (Immediately above her visible right ear, the upper strands sweep diagonally back and up toward that small fold, disappearing behind the far curve of her head.).
| Weights & Prompt | Right-Profile Result |
|---|---|
| No character LoRA + 3 contour sentences | Long hanging ponytail disappeared on all 3; seed 73251 showed a tiny bump behind the crown without hair sweeping toward it |
| v7 + 3 contour sentences | All 3 became ordinary bobs reaching the nape, with only thin stray strands or loops behind the crown rather than a folded bundle |
| v7 + 3 contour sentences + hair-sweep sentence | No long ponytail appeared, but seed 73251 drew a braided-like line above the ear instead of a smooth fan sweeping to the far tie |
Ordered left to right by seeds 73251, 50173, and 89014 below, swapping just the hairstyle prose swung the same 500-step weights from an exposed long ponytail to a plain bob without a visible bundle—still missing 0034’s diagonal sweep into a subtle silhouette crest above a continuous nape.



Testing 13 Right-Profile Prompts on Plain MOSS-Image Base
Every “no character LoRA” test up to this point still had the 4-step distillation LoRA loaded. To see how plain MOSS-Image Base handles right-profile side ponytails on its own, I unloaded both LoRAs and ran Base directly (768×768, Euler 30 steps, CFG 3, flow shift 5, empty negative prompt, no character trigger). Source image 0034 on the left of each strip is strictly a visual reference for humans and was not passed to the model.
Starting at seed 73251, I generated 12 prompts: 2 Illustrious-style tag lists, 4 Anima-style tag-plus-prose hybrids, and 6 prose-only prompts. Here, “Illustrious-style” and “Anima-style” refer to prompt formatting conventions from other anime image models (tag-only lists vs. tags combined with prose), not different model weights—every run used the same MOSS-Image Base with wording varied per prompt. All 12 turned the nose and ear right, yet 9 failed on the back of the head and were dropped without multi-seed comparison grids:
| Dropped Trials (9 Prompts, Seed 73251) | Observed Failure |
|---|---|
Tag-only (2 prompts, incl. left side ponytail, one side up, half updo) | Thick ponytail hanging straight down the back or separated strands on the near side |
| Anima-style tags + prose (4 prompts) | Prompts specifying a small rounded bump produced sideways buns or distorted silhouettes |
| Prose-only (3 dropped prompts) | Either erased the bundle entirely or missed the silhouette shape |
I ran the remaining 3 prose prompts across seeds 50173 and 89014 as well. For the prose prompt describing lower hair overlapping the far bundle’s tip (N-lapped in the image headers), I also created A-matched by prepending and appending standard anime quality/framing tags (such as masterpiece, best quality, score_7, safe) in the exact positions of Kana’s earlier Anima-era captions without altering a word of the prose. Including the 9 single-seed dropped prompts and 4 multi-seed prompts across 3 seeds (9×1 + 4×3 = 21 images across 13 prompts), below are the 4 multi-seed strips with 0034 on the far left followed by seeds 73251, 50173, and 89014.
| Image Label | Prompt Design | Far Bundle & Nape Behavior |
|---|---|---|
N-lapped | Prose only; loose lower hair overlaps far bundle | 73251 kept hair flow and nape, but hung the bundle from the back of the head rather than her left side; 50173 rounded the back; 89014 grew long back hair |
A-matched | N-lapped prose wrapped in Anima-style tags | 50173 peeked only a tiny tuft, yet the skull shape differed from 0034; 73251 and 89014 showed an exposed hanging tail |
N-fan | Prose only; specifies fan-like hair sweep above ear | All 3 showed lower hair and upward flow, though 50173 gathered hair horizontally at ear level |
N-overlap | Prose only; specifies two hair layers and occlusion | 73251 and 89014 hung a short separate tail; 50173 formed a curled ringlet under the ear |




Full prompts for the shared-prose pair (N-lapped & A-matched: right-facing profile tying chestnut hair high above the far left ear, sweeping near-side hair diagonally to the fold while thick lower hair covers the tail)
Prose only (N-lapped, shared across all 3 seeds)
Right-facing anime profile of a brown-eyed girl wearing a white T-shirt on pale grey. She wears her chestnut hair half up, gathered high above her left ear, which is hidden on the far side. The lower layer hangs thick and loose from behind her visible right ear to the base of her neck. Only the rounded top edge of the far-side gathering rises into view, immediately below the crown. On this near side you can trace the upper hair in sweeping diagonal bands from the temple back and up to that little crest. The loose underlying hair covers the hanging length of the gathered section. Bangs and a single curled ahoge.
Anima-style (A-matched, shared across all 3 seeds)
masterpiece, best quality, score_7, safe, 1girl, solo, Right-facing anime profile of a brown-eyed girl wearing a white T-shirt on pale grey. She wears her chestnut hair half up, gathered high above her left ear, which is hidden on the far side. The lower layer hangs thick and loose from behind her visible right ear to the base of her neck. Only the rounded top edge of the far-side gathering rises into view, immediately below the crown. On this near side you can trace the upper hair in sweeping diagonal bands from the temple back and up to that little crest. The loose underlying hair covers the hanging length of the gathered section. Bangs and a single curled ahoge. one side up, half updo, ahoge, brown hair, brown eyes, bangs, profile, from side, facing right, portrait, white t-shirt, light grey background, simple background
All 13 long prompts—at least 39 space-separated words for tag lists and 103 words for prose—failed to reproduce 0034’s head shape and hidden tie across 21 renders. Thinking the overly long phrasing might be pulling other features off-target, I tried stripping prompts down to 6 words.
Nine Short Prompts Starting from Six Words
Keeping the exact same plain Base settings, I ran 9 minimal prompts (01 through 09) at seed 73251, plus an extra seed 50173 on condition 09 for 10 images total.
| Condition | Exact Input Prompt | Outcome |
|---|---|---|
| 01 (seed 73251) | Anime girl with a side ponytail. | Without right-facing phrasing, rendered a three-quarter front view with a long side bundle |
| 02 (seed 73251) | Right-facing profile of an anime girl with a side ponytail. | Nose pointed right, but the tie and tail appeared visibly while pulling the nape up |
| 03 (seed 73251) | Right-facing profile of an anime girl with a left side ponytail. | Turned right, yet adding left still placed a thick bundle on the near side |
| 04 (seed 73251) | Right-facing profile of an anime girl with a side ponytail above her far ear. | Even specifying the far ear left the entire bundle visible |
| 05 (seed 73251) | Right-facing profile of an anime girl with a bob and a side ponytail above her far ear. | Shortened the back hair, but failed to cover the nape like 0034 |
| 06 (seed 73251) | Right-facing profile of an anime girl with a bob and a side ponytail above her far ear. Only its top peeks out. | Looked like an ordinary bob with no distinguishable far-side top edge |
| 07 (seed 73251) | anime girl, profile, facing right, short hair, one side up | Faced right, but hung a long pink ponytail on the near side |
| 08 (seed 73251) | Right-facing anime profile, short bob, half up on the far side. | Produced a horizontal twist behind the ear, too low to match a crown-level far tie |
| 09 (seed 73251) | Anime girl in right profile. Short bob, small left side ponytail just below the crown on the far side. | Short bob with small hair tips sticking straight off the crown, not matching 0034 |
| 09 (seed 50173) | Same as above | Small tuft emerged at the back of the head, missing left-temple placement and sweep |
From left to right: 0034 plus conditions 01–03 (seed 73251), followed by 04–06, 07–09, and finally 09 at seed 50173. Even short prompts either hung the ponytail on the near side, collapsed into a plain bob, or placed tips in the wrong spot.




Testing Built-in Character Names with Left-Side Ponytails
When even minimal prompts could not align the side ponytail, I tested whether existing anime characters known for wearing a left-side ponytail hold their hair-tie placement in right profiles on plain Base without writing any hairstyle words at all. I chose three characters whose official art confirms a left-side ponytail in front view: Mirai Kasuga, Nijika Ijichi, and Fate/Grand Order’s female protagonist Gudako (the common fan nickname for female Ritsuka Fujimaru).
Using the same plain Base settings and seed 73251, I prefixed character and series names with Front-facing anime portrait of or Right-facing anime profile of (writing Gudako (female Ritsuka Fujimaru) from Fate/Grand Order to avoid the male protagonist). Because long ponytails like Nijika’s remain visible behind the silhouette even when occluded, I checked whether the root and hair accessory rotated behind the head or overlapped the near-side hair.
| Character | Front-View Hairstyle | Front vs. Right-Profile Output (Seed 73251) |
|---|---|---|
| Mirai Kasuga | Brown hair with a black hair tie on her left and loose nape hair | Front placed the tie on viewer’s right, but the right profile shifted hair reddish and pulled everything into a high ponytail, changing both face and fastener |
| Nijika Ijichi | Long blonde side ponytail on her left | Front put the long tail on viewer’s right, and the profile kept her triangular ahoge and hair color; however, folds along the back contour made it unclear whether the tie stayed on her left or moved to the back |
| Gudako (Female Ritsuka Fujimaru) | Orange hair tied with a yellow scrunchie on her left | Front showed the yellow scrunchie on viewer’s right; the right profile kept her face, hair color, and yellow scrunchie, yet the thick bundle overlapped the near-side hair instead of tucking behind her head |
Each pair below places the front view on the left and the right profile on the right, in the same order as the table.



Even for Gudako—who kept her signature yellow scrunchie in both views—turning to the right profile pulled the thick bundle across the near side of her head. Across existing characters with built-in left-side ponytails as well, right profiles did not reliably rotate the hair tie behind the far side of the head.