0ppxnhximxr/snapref-v2a

SnapRef v2a

SnapRef v2a stable. Character reference, several views of one character, and per-reference weights. See the examples and v1a / v2a evaluation. Also available on Civitai, with the node and workflows ZIP.

Please support SnapRef! Developing these models has taken 400+ H100 GPU-hours so far. If SnapRef is useful to you, a ⭐ on GitHub, sharing your results or reporting issues helps a lot, and GPU sponsorship is very welcome.

Support SnapRef on Ko-fi

Give Anima one picture of a character. Write only the scene.
Hair, eyes and outfit come from the reference and follow the new pose. The art style comes from the reference too,
or, with mode char, from the base model and your prompt.

Download on Hugging Face · Download on Civitai · Quick start · Character styles · Multi-reference · Reference weights · Evaluation

_{Four references, three scene prompts. Orange outlines mark the inputs. The prompts describe the scene, pose and lighting without character names or appearance tags.}

What you can make

	How to use it
🎭 Character and scene	Use Character & Style (cs) to carry the reference's appearance and visual treatment into another setting.
🎨 Character, your choice of style	Use Character (char) and add watercolor, retro anime or flat-color tags to the prompt.
🧩 Several views of one character	Combine a portrait with a full-body picture to provide more outfit and silhouette information.
⚖️ Per-reference weights	Adjust each picture's contribution. Compare second-reference weights of 0.4 and 1.0 below.

The simple SnapRef node accepts up to eight pictures of the same character. SnapRef (Advanced) exposes the reference layout, individual weights and additional controls.

Character styles

Keep the reference, change the style prompt. These examples use char mode with watercolor, 1990s anime and flat-color treatments; the scene and seed stay the same within each row.

_{The reference supplies the character, while the style prompt changes the rendering. Fine details such as eyewear and collars can still vary.}

Multi-reference

Use two pictures of the same character: one with a clear face and another showing the full outfit. In separate layout, the references are encoded individually and used together during generation.

_{One reference versus two, with the same prompt and seed in each row. The second-reference weight is 1.0 for Persica and 0.4 for Mia. These are selected examples, not an aggregate improvement score.}

The same pair, another scene

The same two pictures can guide a walking pose outdoors or a seated portrait in a study. Keep solo in the prompt when you want one character.

_{Both references contribute to every output. These scenes use independently selected seeds; the controlled one-versus-two comparison is the figure above.}

Reference weights

Start with both weights at 1.0. If a second picture changes the look too much, try lowering its weight. 0.4 gives it a smaller contribution; 0 skips it. Per-reference weights apply in separate layout.

_{Each row keeps the prompt and seed fixed. The first-reference weight remains 1.0. Higher weight is not a guarantee of a better result.}

More scenes

Persica, from a single reference

A museum atrium, a railway platform and a garden bench: different framing and lighting from one input.

After the rain

Three independent characters under the same scene prompt. Cool platform lighting and warm reflections give the reference designs a different atmosphere.

_{Selected examples generated with SnapRef v2a.}

Quick start

1. Install Anima base v1.0 with its text encoder and VAE in ComfyUI.
2. Copy comfyui/snapref/ from this repository into ComfyUI/custom_nodes/snapref/, then restart ComfyUI.
3. Download snapref_v2a.safetensors to ComfyUI/models/loras/.
4. Open workflows/snapref_v2a_simple.json, choose your own reference image in the SnapRef node and queue. The adapter is already set to snapref_v2a.safetensors.

Optional FP8 download: snapref_v2a_fp8.safetensors is approximately 195 MB (about 49% smaller). Put it in the same models/loras/ folder and select it in the SnapRef adapter field. It uses E4M3FN matrices with per-row scales; the node restores them to BF16 for inference. Generation results can differ slightly from the original weights.

Use an up-to-date ComfyUI installation. If your base-model, text-encoder or VAE filenames differ, select the installed files in the loader nodes. In the advanced workflows, choose a picture in each Load Image node.

Try a scene-only prompt like this:

masterpiece, best quality, score_7, safe, 1girl, solo, sitting on a wooden garden bench,
looking at viewer, gentle smile, dappled sunlight, flowering trees

Choose the mode before writing style tags. Use Character & Style for the reference's look, or Character to try another rendering style. In the simple node, fill two picture slots and click + beside Character to add more views.

SnapRef uses its own node. Loading these weights through LoraLoader or a generic Diffusers pipeline is not supported.

Workflows

Workflow	Use
Simple	Pick images inside the SnapRef node; Character & Style, Strength and Fidelity.
Single reference	SnapRef (Advanced) with a separate image input.
Character mode	Take the character from the reference and set the style in the prompt.
Multi-reference	Two views of the same character.
Reference weights	Two references with the second weight set to 0.4.
Two characters	Stitch references side by side and describe two characters in the prompt.

Settings

The showcased images use CFG 4.5 · Euler / simple · Fidelity 1.0 on Anima base v1.0.

Simple SnapRef control	Starting point	Purpose
Adapter	snapref_v2a.safetensors	Select the v2a stable weights.
Mode	Character & Style	Use Character when the prompt should set the rendering style.
Strength per picture	1.0	Contribution of that reference; 0 disables it.
Fidelity	1.0	Lower values allow more prompt freedom.

SnapRef (Advanced) settings

Input	Starting value	Purpose
mode	cs	cs: character and style; char: character with prompt-driven styling.
strength	0.0	Global reference attention on a log scale; different from the simple node's per-picture Strength.
fidelity	1.0	Ignored when ref_in_uncond or ref_guidance_on is enabled.
ref_megapixels	1.0	Maximum area of each encoded reference; smaller images are not enlarged.
ref_caption	empty	Optional reference appearance tags.
ref_in_uncond	off	Keep the reference out of the negative branch at the showcased default.
ref_guidance_on	off	Additional reference guidance; the figures use it disabled.
image_2, image_3	optional	Additional images of the same character.
use_batch	off	Read multiple images from an IMAGE batch.
multi_layout	separate	Encode references independently. stitch combines them into one image.
weight_1 … weight_3	1.0	Per-reference contribution in separate layout.

How it compares

The fixed evaluation contains 64 cases × 2 seeds, separate from the curated figures above.

Metric	v1a	v2a
Appearance tag recall ↑	0.8722	0.8727
Extra appearance tags ↓	0.1341	0.1401
Head SigLIP similarity ↑	0.8989	0.8903
Head DINO similarity ↑	0.7431	0.7407
Head CCIP distance ↓	0.0684	0.0715
Full-image CCIP distance ↓	0.0725	0.0757

Appearance-tag recall improves slightly, while our qualitative assessment shows substantially better reference following than v1a. A remaining issue is that the model can struggle to reproduce facial details when the face in the reference image is very bright.

The metrics are sensitive to framing, pose, lighting and head detection.

Things to watch for

• Small accessories, collar shapes and fine patterns can change, especially with a strong style prompt.
• An upper-body reference does not define unseen lower-body clothing; a second full-body view can help.
• Artwork text or signature-like marks can be copied from a reference. Inspect outputs before sharing them.

Release details

Release	SnapRef v2a stable
Base used for these examples	Anima base v1.0
Weight file	snapref_v2a.safetensors
File size	381,821,960 bytes, approximately 382 MB

Files

File or folder	Contents
snapref_v2a.safetensors	v2a stable adapter weights.
snapref_v2a_fp8.safetensors	Optional FP8 storage variant, 194,652,680 bytes; inference computes in BF16.
comfyui/snapref/	SnapRef and SnapRef (Advanced), including the image-picker interface.
workflows/	Six ComfyUI workflows configured for v2a.
figures/	Release figures and hero preview.
SHA256SUMS	Checksums for the weights and release files.
LICENSE.md · NOTICE.txt	Weight license and attribution.

License

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