Comparing SnapRef v1b with Ref Adapter v2 and Role-Ref-KV on M1 Max Under Identical Conditions

Right before publishing my previous article comparing Anima Ref Adapter v2 and Role-Ref-KV, another adapter serving the same purpose was released.
It is SnapRef, designed to maintain a character’s features across different compositions and scenes from a single reference image.
The comparison table in its README reports numbers showing it is faster than Ref Adapter v2 and Role-Ref-KV, and closer to the ground truth image than v2.

However, those numbers were measured on an A100.
Numbers alone do not tell us what happens when tested across the same 8 conditions as before, or whether the features unique to SnapRef’s README function identically under Mac MPS.
Therefore, using the exact same reference images, prompts, and seeds as the previous article, I tested v1b using ComfyUI on an M1 Max 64GB setup.

Overview of SnapRef

The underlying mechanism belongs to the same family as Ref Adapter v2 and Role-Ref-KV.
The reference image passes through the model once via a dedicated path, caching the key/value representations from the attention mechanism across each block, which are then read during each step of generation.
Since it wraps the model in ComfyUI, standard samplers can be used directly.

The differences between v1a and v1b are summarized below.

The README describes v1b as having “double the rank” and recommends starting with v1a if undecided.
Examining the weight keys reveals that rank is not the only difference.
LoRA modules number 336 in v1a versus 392 in v1b, with v1b also applying to the mlp layers on the generation side.

The comparison table in the README also lists values for other adapters.

The ground truth image is an average over 52 cases where the same character adopts a specified pose in a separate image.
Looking into the distributed scores_v1a_v1b.json, DINOv2 similarity directly against the reference image itself measures 0.552 for v2, 0.515 for v1a, and 0.510 for v1b, reversing the order.
The ranking flips depending on which metric is examined.

Recommended settings from the README:

Test Environment

ref_guidance_on does not run out of the box on MPS.
It halted with Cannot convert a MPS Tensor to float64 dtype, so just as with Ref Adapter previously, I changed two lines of APG correction from .double() to .float().
The remainder of the node code was left intact, except for a temporary one-off test in the black image section.

The tested models covered the same 8 conditions as before, matching the recommended 30 steps and cfg 4.5 on Base.

The 2.9B remapping reindexes key numbers to skip the 12 interleaved blocks out of 40.
Since the node only applies LoRA to modules with existing keys, zero-padding like that required for Role-Ref-KV was unnecessary.

By default, SnapRef reads the reference only on the positive prompt side.
At cfg 1.0 for Turbo, model evaluation occurs once per step, whereas at cfg 4.5 it occurs twice.

Running with Official Settings

Tested on Base using the README prompt, workflow negative, 832×1216 resolution, 28 steps, and cfg 4.5.

Kana-chan

ref_girl

masterpiece, best quality, score_7, safe, 1girl, solo, full body, sitting, crossed legs, bench, street, night

worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration

I also ran the cake cafe scene with ref_girl under identical conditions to the previous article.
The output showed her bringing a fork to her mouth on a red sofa, preserving hair color, streaks, and clothing, taking 302.6 seconds.

SnapRef v1b

Ref Adapter v2 (Previous)

Role-Ref-KV (Previous)

Results Across Base, Aesthetic v1.1, and 2.9B

Unless specified otherwise, subsequent sections use Kana-chan’s standing reference and the negative prompt from the previous article.

Base Running

Base Cake

Aesthetic v1.1 Running

Aesthetic v1.1 Cake

2.9B (remapped) Running

2.9B (remapped) Cake

2.9B (unmapped) Running

2.9B (unmapped) Cake

For the cake scene, all three functional conditions produced the cowlick, side ponytail on the right, white shirt, and red tie faithful to the reference.
The slight orange tint in the hair observed with Role-Ref-KV in the previous cake test did not appear here.

2.9B failed to function without remapping.
Rather than collapsing into a grid pattern like v2 did, the figure remained sharp in the original standing pose while the prompt’s scene was scarcely reflected.

In the running scene, Base and Aesthetic faced backward.
Since the prompt lacked directional constraints, this was not inherently incorrect, but since v2 and Role-Ref-KV previously faced forward with the same seed, I investigated the orientation further in a later section.

masterpiece, best quality, score_7, newest, highres, absurdres, safe, 1girl, solo, full body, running, wind, hair blowing, dynamic pose, fantasy landscape, castle in background, sunset sky, dramatic clouds, grass field

masterpiece, best quality, score_7, newest, highres, absurdres, safe, 1girl, solo, upper body, sitting, eating, cake, cafe, indoors

worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, chromatic aberration, bad anatomy, bad hands, extra arms, extra fingers, missing fingers, text, watermark, signature, artist name, multiple views, split screen, comic, 2koma, 4koma

Results on Turbo Models

Turbo v1.1 Running

Turbo v1.1 Cake

Turbo v1.0 Running

Turbo v1.0 Cake

Base + Turbo LoRA Running

Base + Turbo LoRA Cake

2.9B unofficial Turbo Running

2.9B unofficial Turbo Cake

All four Turbo conditions faced forward in the running scene, avoiding the backward orientation seen on Base.
While the README targets Anima base v1.0, SnapRef operated seamlessly across full Turbo models, Turbo LoRA, and remapped 2.9B alike.
The red tie being occluded by an arm in the cake scene was an artifact of composition and remained visible under settings explored in the next section.

Testing 4-Panel Collapse with ref_guidance 4.5

In earlier testing, the pilot adapter split into 4-panel comic layouts when run at cfg 1.0 and ref_guidance 4.5 on Turbo.
Ref Adapter v2 avoided 4-panel splits but suffered desaturated colors.
Since SnapRef provides ref_guidance_on, I generated the cake scene under identical settings.

Turbo v1.1

Turbo v1.0

Base + Turbo LoRA

2.9B unofficial Turbo

None of the four conditions split into multiple panels.
Evaluating the model twice per step increased runtime by roughly 1.5x, though with a single seed it is difficult to separate deliberate effect from random variance.

I also tested ref_guidance_on on Base.
The README notes that values of 6–8 increase reference adherence, taking ~1.4x runtime, and that ref_end 0.6 provides faster execution with nearly identical visual results.

ref_guidance 4.5

ref_guidance 7

ref_guidance 7, ref_end 0.6

Turbo v1.1, ref_guidance 2

Measured times on M1 Max were roughly 1.5x of baseline with guidance disabled, and just under 1.4x even with ref_end 0.6.
The claim that ref_end 0.6 yields nearly identical images held true in practice.

Testing Lying and Overhead Compositions

Turbo v1.1

Base

From a single frontal standing reference, the character features were preserved cleanly across both supine and overhead perspective shifts.

masterpiece, best quality, score_7, newest, highres, absurdres, safe, 1girl, solo, full body, lying on back, on grass, from above, looking at viewer, flowers, daytime

Evaluating Seed Variation in the Running Scene

Adding seeds 20261028 and 20261029, I tested across the 4 Turbo setups alongside Base and remapped 2.9B.

Turbo v1.1 s20261027

Turbo v1.1 s20261028

Turbo v1.1 s20261029

Turbo v1.0 s20261027

Turbo v1.0 s20261028

Turbo v1.0 s20261029

Base + Turbo LoRA s20261027

Base + Turbo LoRA s20261028

Base + Turbo LoRA s20261029

2.9B unofficial Turbo s20261027

2.9B unofficial Turbo s20261028

2.9B unofficial Turbo s20261029

Base s20261027

Base s20261028

Base s20261029

2.9B remapped s20261027

2.9B remapped s20261028

2.9B remapped s20261029

Across all 18 generations, visible hairstyle and clothing traits strictly followed the reference.
In the 14 forward-facing images, the side ponytail, white shirt, red tie, and black tights were consistently rendered. The 4 backward-facing outputs also retained the scrunchie, ponytail, and tights.
Unlike Role-Ref-KV previously, there were no seeds where the ponytail or tie disappeared.
Conversely, Base faced away across all 3 seeds, whereas v2 and Role-Ref-KV had both faced forward under these exact seeds.

Isolating Why Base Runs Facing Away

To determine whether facing away was an idiosyncrasy of v1b or Base itself, additional seeds and conditions were tested.

No Reference s20261030

v1b s20261030

v1a s20261030

Ref Adapter v2 s20261030

Role-Ref-KV s20261030

No Reference s20261031

Base without any reference faced backward on both seeds, indicating that facing backward under this prompt is an inherent bias of Base itself.
Unlike v2 or Role-Ref-KV, which are strongly pulled toward the frontal composition of the reference illustration, v1b did not overwrite the model’s compositional bias.
This aligns with the README’s description that SnapRef “extracts appearance rather than copying composition,” though the sample size remains limited.

Increasing reference strength did not alter orientation, whereas adding looking at viewer to the prompt immediately restored a frontal view.

strength 1

ref_guidance 7

looking at viewer added

Comparing v1a, v1b, and FP8

v1b Turbo v1.1 Running

v1a Turbo v1.1 Running

v1b FP8 Turbo v1.1 Running

v1b s20261028

v1a s20261028

v1a s20261029

v1b Turbo v1.1 Cake

v1a Turbo v1.1 Cake

v1b FP8 Turbo v1.1 Cake

v1b Base Running

v1a Base Running

v1b FP8 Base Running

Across the 3 seeds of the Turbo v1.1 running scene, v1a exhibited clothing drift twice, compared to 0 for v1b.
While README reported similar ground truth similarity metrics, in this reference and seed set, v1b retained clothing details more faithfully.

FP8 halved file size and produced “virtually identical images” as claimed in the README.
Pixel-level absolute difference against bf16 averaged 1.8–3.0 on a 0–255 scale.
The composition remained unchanged, differing only subtly in fine detail.
Generation speed was 56.5–58.6 s (Turbo) and around 312 s (Base), practically identical to bf16.
Since ComfyUI was launched with --bf16-unet, the exact precision used internally during FP8 computation was not examined.

Effects of the strength Parameter

strength in the Advanced node operates on a logarithmic scale, altering key weights from the reference.
The tooltip indicates that +1 multiplies influence by ~2.7x, while -1 divides by ~2.7x.
The README suggests values between 1 and 2 when seeking to enhance fine details (accessories, sleeves, decorations).

Turbo Running -1

Turbo Running 0

Turbo Running 1

Turbo Running 2

Turbo Cake -1

Turbo Cake 0

Turbo Cake 1

Turbo Cake 2

Base Cake -1

Base Cake 0

Base Cake 1

Base Cake 2

As strength increased, the pose converged toward the standing reference, dampening running dynamics and expressive variety.
At 2 on Base, colors became washed out.
Because Kana-chan’s illustration lacks intricate adornments, whether fine details actually increase could not be established with this reference.

Adjusting strength did not increase runtime.
Turbo v1.1 cake took 56.5 s, running took 45.4–45.5 s, and Base cake took 301.8–302.0 s, showing no penalty compared to 0.
Although the README states that “strength and weighting no longer slow things down,” the tooltip still bears an outdated “~1.3x slower” note.
Even though MPS falls back to a masked path without fast fused kernels, execution speed was unaffected.

Note that the simpler SnapRef node uses a different scale for strength: 1 corresponds to training default, 0 is off, and 2 is 2x, without logarithmic scaling.
A value of 1.0 in SnapRef produced an identical image to 0 in Advanced.
The README’s recommendation to “try 1 to 2” applies to the Advanced node; a setting of 2 in simple SnapRef corresponds to roughly 0.7 in Advanced.

fidelity Behaves as a Binary Toggle on Mac

fidelity lowers reference adherence in favor of prompt compliance. Inspecting the node’s code reveals distinct execution between CUDA and non-CUDA environments.
On CUDA, the negative prompt side attends to the reference weighted by 1 - fidelity, relying on optimized kernels.
On non-CUDA systems, any value below 1 acts as a fallback where the negative side fully attends to the reference.

Testing on the Base cake scene confirmed this behavior:

fidelity 1.0

fidelity 0.5 (0.0 and ref_in_uncond yield identical output)

In line with the implementation, setting 0.5 or 0.0 on Mac produced the exact same output as enabling ref_in_uncond, with no intermediate gradients.
Turbo at cfg 1 does not compute the unconditional side, so fidelity settings produced no effect.

Changing Art Style with mode: char

mode: cs extracts both character and art style from the reference by default, while char aims to extract character only.
The README highlights pairing char with artist tags.
Here, I tested general style tags rather than specific artists.

Base watercolor cs

Base watercolor char

Base 1990s cs

Base 1990s char

Under cs, adding style tags barely altered the visual output.
Under char, stylistic tags took effect, though watercolor altered character hair color rather than producing a watercolor texture.
The “What’s next” section of the README notes plans to strengthen mode char in v2a.

Generating Two Characters from a Single Stitched Reference

Following the README’s guidance, I stitched two full-body character illustrations side by side into a single reference image, generating at 1216×832.

Reference (1664×1216)

Turbo v1.1 s20261027

Turbo v1.1 s20261028

Turbo v1.1 s20261029

Base s20261027

Base s20261028

All 5 outputs successfully produced both characters without swapping horizontal placement or hair colors, preserving black tights and white socks.
However, Kana-chan’s necktie turned into a ribbon matching the right character in 4 of 5 outputs.
On Base, the right character also acquired Kana-chan’s side ponytail.
While the README mentions that “attributes not clearly discernible in the reference may bleed between characters,” attribute mixing occurred even for distinct elements like ties and skirts.

masterpiece, best quality, score_7, newest, highres, absurdres, safe, 2girls, brown hair girl on the left, blonde girl on the right, standing side by side, full body, looking at viewer, park, day

Adjusting strength on Stitched References Produces Black Images

Attempting to curb attribute bleeding by raising strength to 2 unexpectedly produced completely black images.
Varying the seed, switching to Base, or setting strength to 1 or -1 yielded identical black results.
Conversely, the same reference image generated normally when strength was left at 0.

Isolating the trigger revealed that having two characters was irrelevant; the failure hinged strictly on the reference image resolution.
The node scales reference images to ~1MP or less in multiples of 16.
A 1664×1216 image downscales to 1184×864.
A single Kana-chan illustration padded to that same resolution also failed to render, whereas other aspect ratios succeeded.

Each 16 pixels correspond to 1 token.
The _masked_attention path on MPS pads key counts to multiples of 16, setting the padded mask region to float("-inf").
At strength 0, this masked path is bypassed, preventing black images.
Changing that mask value from -inf to -1e4 and re-running with an 1184×864 reference at strength 2 (seed 20261027) resolved the black image issue.
Although top and bottom white padding bands appeared, the tie, ribbon, plain skirt, and plaid skirt rendered faithfully for both characters.
However, passing identically sized random tensors through PyTorch attention alone did not trigger NaN, so the exact underlying cause of the failure under -inf remains unisolated.
The code modification was reverted after testing.

Node with mask changed to -1e4, strength 2

To avoid this issue without editing node source code, simply resize the reference so that the total token count forms a multiple of 16.
Padding the two-character reference to 1216×832 allowed generation to proceed cleanly at both strength 1 and 2.

strength 0

strength 1

strength 2

strength 2, s20261028

Preserving the tie succeeded in only one image at strength 1 and reverted at strength 2, indicating that adjusting strength is not a reliable safeguard against attribute bleeding.
The black image mention in the README refers to an fp16-related bug, which is distinct from this phenomenon.
Since CUDA takes an alternate fused execution path, this black image appears specific to environments like MPS where fused attention kernels are unavailable.
Adjusting strength away from 1 in the simple SnapRef node or modifying multi-reference weight routes through the same code path and may trigger identical failures depending on reference dimensions.

Using Multiple References for the Same Character

A node update on November 1 introduced support for up to 3 references for a single character.
The README highlights pairing a bust-up reference with a full-body reference.
I tested this using Kana-chan’s bust-up portrait alongside her full-body standing illustration.
The bust-up portrait features a blazer and ribbon, contrasting with the full-body’s white shirt and tie, making it straightforward to identify which reference influenced the output.

Bust-up Reference

Full-body Reference

Bust-up only

separate

stitch

Bust weight 0.4

Bust-up only Cake

separate Cake

Full-body weight 0.4 Cake

Full-body 1st Cake

Adding the full-body reference successfully completed the lower-body attire (plain skirt, black tights, loafers), which had hallucinated unreferenced clothes when using the bust-up image alone.
The approach advocated by the README functioned as advertised.
All configurations generated a single character without duplicating into two figures.

When combining two references with conflicting outfits, the blazer from the bust-up image took precedence across both scenes when weights were equal.
Reversing the slot order yielded identical dominance, requiring the bust-up image weight to be reduced to 0.4 before the full-body shirt and tie emerged.
Images occupying a larger relative proportion of the character’s facial area may exert stronger attention pull.
On Base, separate mode required 350.4 s, roughly 40 seconds longer than a single reference.

Combining with 3D Poses

In previous work, pairing Ref Adapter with LLLite required routing through CharRef-LLLite.
Because SnapRef maintains separate reference key/value caches without interleaving reference tokens into the generation-side sequence, stock nodes seemed likely to work.
I compared ComfyUI’s standard AnimaLLLiteApply against CharRef-LLLite’s ExperimentalAnimaLLLiteRefApply.
Pose input used the same 3dpose_example.png at 768×1152 with an LLLite strength of 1.0.

Pose Input

SnapRef Reference Only

Stock LLLite s20261027

Stock LLLite s20261028

Stock LLLite s20261029

Stock LLLite Base

Stock nodes reproduced the target pose accurately, matching CharRef-LLLite pixel-for-pixel across all 3 seeds.
SnapRef eliminates the need for specialized adapter bridge nodes.
Generation time on Turbo was 45.3–45.4 s, noticeably faster than the 86.4 s required previously with v2.
Base (30 steps, cfg 4.5) also adhered strictly to the pose in 272.5 s, rendering loafers cleanly.
The missing shoes occurred only on Turbo seed 20261027, which had similarly lost shoes in the previous v2 evaluation.

masterpiece, best quality, score_7, newest, highres, absurdres, safe, 1girl, solo, full body, sitting, looking at viewer, simple background

Testing Known Limitations from the README

The README highlights two specific limitations: scene prompts overriding attire (such as beach forcing swimwear) and outdoor full-body shots rendering white floors underfoot, which supposedly resolves by explicitly prompting ground textures.

beach Base

beach Base strength 2

beach Turbo v1.1

beach Turbo v1.1 strength 2

outdoors Base

outdoors, grass Base

Across 4 generations with a single seed, swimwear never appeared.
The shift from long to short sleeves on Base represented the extent of clothing override.
Adding grass created an isolated circular patch of lawn beneath the character rather than rendering an integrated terrain.
The README’s claim that prompting ground textures resolves floor artifacts did not hold true in this instance.

Evaluating simplification of minor details or brightening on dark-skinned references was omitted due to the lack of suitable reference assets.

masterpiece, best quality, score_7, newest, highres, absurdres, safe, 1girl, solo, full body, standing, beach, ocean, day

masterpiece, best quality, score_7, newest, highres, absurdres, safe, 1girl, solo, full body, standing, looking at viewer, outdoors, day

Generation Speed Benchmarks

Measured generation times for 896×1152 images on M1 Max 64GB.
Tests were run consecutively on the same day in the running scene across rotating seeds.

Additional benchmark figures:

The reference overhead on Base was approximately 50 seconds, or ~1.2x of the unguided baseline.
This aligns closely with the ~1.3x overhead reported on A100 in the README (9.9 s vs. 7.7 s).
On Base, v1b required roughly two-thirds the time of v2, matching the ratio in the README (9.9 s vs. 14.4 s).

Because Role-Ref-KV operates at a recommended 28 steps, its total runtime was comparable to v1b.
Per-step execution averaged ~10.3 s for v1b and ~10.6 s for Role-Ref-KV, showing negligible difference.
While README benchmarks on A100 showed v1b at 9.9 s versus Role-Ref-KV at 12.5 s, that margin did not materialize on Mac.
On Turbo, however, because Role-Ref-KV evaluates twice per step even at cfg 1, v1b was roughly 1.7x faster.
Performance relative to v2 was virtually identical.

The difference between v1a and v1b on Turbo was around 1 second, within the margin of seed variation.
On Base, v1a clocked in at 298.0–308.3 s, running only ~1–5% faster than v1b.

Unexamined Scenarios

• The fp16 black image issue marked as resolved in the README (launching with --bf16-unet bypasses reproduction conditions)
• Combinations with Anzhc/sorryhyun Mod Guidance or Spectrum
• FP8 quantized Anima base
• Adjusting ref_caption, use_batch, correction’s rescale, or ref_megapixels
• General CUDA-specific behavior (including intermediate fidelity steps and black image constraints)
• Photorealistic, dark-skinned, or heavily accessorized reference images
• Peak VRAM footprint comparisons

Seed counts ranged between 1 and 5 per condition, focused on Kana-chan and ref_girl.
Phrases indicating that an issue “did not occur” are bound strictly to this evaluation scope.

Practical Recommendations

Reference Links

• snapref-v1b (Hugging Face)
• snapref-v1a (Hugging Face)
• SnapRef (GitHub)
• ComfyUI-Anima-Ref-Adapter (GitHub)
• Anima-Role-Ref-KV (Hugging Face)
• ComfyUI-Anima-CharRef-LLLite (GitHub)
• Anima_ControlNet-LLLite (Hugging Face)
• Anima (Hugging Face)
• Anima-2.9B (Hugging Face)
• anima29b-turbo-unofficial (Hugging Face)
• Comparing Anima Ref Adapter v2 and Role-Ref-KV
• Testing Base-Dedicated Anima Ref Adapter on Turbo, Aesthetic, and 2.9B