0ppxnhximxr/snapref-v1b

SnapRef v1b

One reference image. Your character, in any pose and scene you prompt.

A reference adapter for Anima base v1.0. The reference is read once and reused at every sampling step, so it adds only a little to generation time.

v1b is the larger sibling of SnapRef v1a: the same method and data, with twice the LoRA rank. It follows the prompt slightly more closely, and keeps the character about as well as v1a. If you are unsure which one to use, start with v1a: it is smaller and a little faster.

	v1a	v1b
LoRA rank (reference / target branch)	128 / 64	256 / 128
file size, bf16 / FP8	382 / 195 MB	910 / 460 MB
seconds per image ¹	8.9	9.9
DINOv2 similarity to the true image ² ↑	0.722	0.720
CLIP image-text similarity (prompt following) ² ↑	0.249	0.256

_{¹ A100, 832×1216, 28 steps, default settings. ² 52 cases with the same seeds, 32 of them characters not seen in training; the "true image" is another picture of the same character in the prompted pose. Per-case scores are in scores_v1a_v1b.json.}

Thanks. This project started from the ideas in ElectronGirl's ComfyUI-Anima-Ref-Adapter. Their early work showed that Anima can read a reference image this way. Thank you.

---

Quick start

1. Install Anima base v1.0 with its text encoder and VAE.
2. Copy comfyui/snapref/ into ComfyUI/custom_nodes/ and restart ComfyUI. It is the same node as in v1a.
3. Put snapref_v1b.safetensors (or snapref_v1b_fp8.safetensors) into ComfyUI/models/loras/.
4. Open workflows/snapref_v1b.json, load a reference image and queue.

Write only the scene. Hair, eyes, outfit and style come from the reference.

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

The adapter is not a regular LoRA. Load it with the SnapRef node, not with LoraLoader.

Settings

setting	default	what it does
mode	cs	cs copies the character and the style. char copies only the character, so you can set the style with artist tags
strength	0	how closely to follow the reference. Try 1–2 for more detail (trims, sleeves, accessories)
ref_guidance_on	off	a stronger push toward the reference, at about 1.4× the time
ref_guidance	4.5	strength of that push; 6–8 for a closer match

All settings

setting	default
adapter	snapref_v1b.safetensors	the bf16 file or the FP8 file
ref_in_uncond	off	on: the reference is followed more loosely and the prompt more closely
ref_megapixels	1.0	the reference is shrunk to this size if it is larger
ref_caption	empty	optional appearance tags of the reference; can help with accessories
correction	apg	with ref_guidance_on: keeps colors from over-saturating
ref_end	1.0	with ref_guidance_on: 0.6 is faster with nearly the same result

Recommended sampling: euler or er_sde, simple, 28–30 steps, cfg 4.5.

Two characters

Put the two references side by side with the core Image Stitch node and connect the result as the reference. In the prompt, use 2girls (or 1boy, 1girl) instead of solo, and say in a few words who is on the left and who is on the right. A landscape size works best. See workflows/snapref_v1b_two_characters.json.

Results

_{Characters that were not in the training data. Each column uses the same prompt and seed. The prompt describes only the scene.}

More examples

	SnapRef v1b	SnapRef v1a	ElectronGirl v2	Role-Ref-KV	Anima, no reference
seconds per image ¹	9.9	8.9	14.4	12.5	7.7
DINOv2 similarity to the true image ² ↑	0.720	0.722	0.675	–	0.493

On characters not seen in training, the adapter predicts images better with the right reference than with a wrong one: identity gap 0.0038 on the held-out denoising loss (v1a: 0.0040). The numbers are in heldout_loss.json.

How it works

The reference goes through the model once, through its own branch, and every layer's attention keys and values are cached. While the image is generated, each layer attends to its own tokens and to the cached reference. Each branch has its own LoRA. The reference sits at a separate position in the model's coordinate space, so it is read for appearance and not copied as a layout.

Training used the same 58,300 Anima-generated images as v1a (scenes3x4), 10,000 steps at batch 16, on an RTX PRO 6000 Blackwell. training_config.json has every setting.

FP8

snapref_v1b_fp8.safetensors is half the size (460 MB) and gives practically the same images. It also works with the FP8 Anima base.

Compatibility and troubleshooting

• Black image with an fp16 model or --fast. Fixed in the node update of 2026-10-31. Anima overflows in fp16. ComfyUI runs the model in fp16 when the checkpoint is stored in fp16 (many merges), with --fp16-unet, or with --fast / --fast fp16_accumulation. The old node encoded the reference in fp16 there and produced NaN. The new node keeps that part in fp32, as ComfyUI itself does. bf16 output is unchanged (same bits). With the old node, add --bf16-unet to the ComfyUI launch arguments, or drop fp16_accumulation from --fast.
• Anzhc Mod Guidance: works in either node order with the updated node. Before, applying Mod Guidance first dropped the reference.
• Spectrum: works with ref_guidance_on off. With it on, Spectrum-type step forecasting breaks; the node now logs a warning.
• sorryhyun Mod Guidance: with the default settings it can stop with a tensor-size error (4 vs 2). Turning on ref_in_uncond or ref_guidance_on avoids it.

Known limitations

• Very small details (earrings, fine patterns, scars) can be simplified.
• A dark-skinned reference can come out with lighter skin.
• With some outdoor full-body prompts, the bottom of the image can turn into a plain white floor. Naming the ground (grass, path) fixes it.
• A scene can override the reference outfit (beach → swimsuit). Raise strength or name the outfit.
• With two characters, attributes the references do not show clearly (for example, closed eyes) can be mixed between the two.

What's next

v2 changes the data, not the size: generated images cut to half or less, real Danbooru-style illustrations added (100–150k images in total), outfit-change pairs and multi-resolution training.

Files

file
snapref_v1b.safetensors	adapter, bf16 (910 MB)
snapref_v1b_fp8.safetensors	adapter, FP8 (460 MB)
comfyui/snapref/	ComfyUI node (same as v1a)
workflows/	single character, two characters, side-by-side comparison with other adapters
scores_v1a_v1b.json	per-case embedding scores of v1a, v1b, ElectronGirl v2 and base

License

The weights are a derivative of the CircleStone Model and are distributed under the CircleStone Labs Non-Commercial License v1.2 (LICENSE.md, NOTICE.txt). Non-commercial use only. Not affiliated with or endorsed by CircleStone Labs.