SnapRef V2A for Anima in ComfyUI: Same Character in Every Scene

Keeping the same anime character across a set of images is one of the hardest things to do with a text prompt alone.

SnapRef v2a is a reference adapter for the Anima image model that fixes that: drop in a picture of your character, write only the new scene, and the hair, eyes and outfit carry over into the new pose.

v2a is the stable release, and it adds the two things people asked for most: a separate weight for each reference picture, and a Character-only mode that lets your prompt control the art style.

This guide covers the files, the install, both nodes, every setting, the six included workflows, multi-reference tips, the honest v1a vs v2a numbers, and fixes.

What SnapRef v2a Is

Anima is an open anime-style text-to-image model from CircleStone Labs. On its own, every image is a fresh roll of the dice: describe “a girl with silver twin tails and a red jacket” twice and you get two slightly different girls.

SnapRef adds a reference image. You give it one or more pictures of your character, and the prompt only has to describe the scene, such as “sitting in a cafe, rainy window, evening light”. The adapter reads the character from the pictures and puts them into the new scene and pose.

The first public release, v1a, appeared at the start of November as a cached-KV reference adapter: the reference is encoded once and its attention keys and values are reused while the image is generated. v2a is the stable follow-up, released on 8 November.

What Is New in v2a

• A weight per reference. In the default separate layout, each picture has its own weight. It defaults to 1.0, and 0 switches that picture off. You can now let one picture lead and another just add detail.
• Two modes in the simple node. Character & Style (cs) takes both the look of the character and the art style from the reference. Character (char) takes only the character, so artist tags and style words in your prompt set the look, such as watercolour or flat colour.
• Six workflows, including a two-character workflow and a reference-weight comparison.
• Better reference following, by the authors’ visual judgement. The automatic metrics are level with v1a, as covered below.

Files and Folders

The FP8 file stores the weight matrices in E4M3FN with per-row scales and turns them back into bf16 when it loads. It is half the download, but results can differ slightly from the full file.

The SnapRef page notes that Anima’s file names can differ between downloads, so just pick whatever you installed in the loader nodes. If you do not have Anima set up yet, our Anima Reference Adapter v2 guide walks through the base model install.

Install

1. Install Anima base v1.0 with its text encoder and VAE.
2. Download the SnapRef v2a repo from Hugging Face.
3. Copy its comfyui/snapref/ folder into ComfyUI/custom_nodes/snapref/.
4. Put snapref_v2a.safetensors in ComfyUI/models/loras/.
5. Restart ComfyUI.
6. Open workflows/snapref_v2a_simple.json and pick your reference image in the SnapRef node.

One rule matters more than any other: load it with the SnapRef node, never with LoraLoader. The file holds two LoRA branches behind one key, and a normal LoRA loader will not read it correctly. A generic Diffusers pipeline will not load it either.

There is no separate GitHub node repo to clone; the node folder ships inside the Hugging Face repo. The training and evaluation code lives at ArcAnima/SnapRef.

The Simple SnapRef Node

This is the one to start with. It takes up to eight pictures of the same character.

Drop a picture onto the image slot inside the node, write a scene-only prompt, and queue.

The SnapRef (Advanced) Node

The advanced node exposes everything. Most people will not need it, but these are the inputs and defaults from the model card:

• mode: cs (default) or char.
• strength: 0.0 by default. A global reference-attention boost on a log scale.
• fidelity: 1.0. Ignored when ref_in_uncond or ref_guidance_on is on.
• ref_megapixels: 1.0. The largest area each reference is encoded at.
• ref_caption: optional appearance tags for the reference.
• ref_in_uncond and ref_guidance_on: both off by default.
• image_2, image_3: optional extra views of the character. use_batch is off by default.
• multi_layout: separate (default, each picture encoded on its own) or stitch (pictures joined into one image).
• weight_1 to weight_3: 1.0 each. These only apply in the separate layout.

The Six Workflows

The simple graph and the showcase images use 28 steps, Euler sampler, simple scheduler, CFG 4.5, fidelity 1.0, at 832 x 1216. If those sampler names are new to you, see our samplers and schedulers guide.

Using Several References

Multiple pictures help when one image does not show everything. A head-and-shoulders picture tells the model nothing about the trousers or shoes, for example.

• Start every weight at 1.0. If the second picture changes the look too much, lower its weight. The two-reference workflow sets the second weight to 0.4.
• Add a full-body view when the main reference is upper-body only.
• Use separate layout when you want per-picture weights. stitch joins them into one image and ignores the weights.
• Keep references consistent. Two pictures with different outfits will blend into something in between.

Prompting Tips

• Write the scene only. Let the reference handle hair, eyes and clothes. Describing them again can fight the reference.
• Choose the mode first. Pick Character & Style or Character before you add style tags. In Character & Style mode, style tags have less effect.
• Add “solo” when you want exactly one character, so the model does not add a twin.
• Lower fidelity a little (for example 0.8) if the pose or expression will not change from the reference.

v1a vs v2a: The Numbers

The authors tested 64 cases with two seeds each, and they are upfront that v2a is “level with v1a rather than ahead of it” on the metrics.

So why upgrade? Their visual read is that v2a follows the reference “substantially better”, and the new controls (per-picture weights and Character mode) are worth having on their own. The metrics also shift with framing, pose and head detection, so small differences like these are close to noise.

SnapRef vs Anima Ref Adapter

Both are good. SnapRef gives you more control when you have several pictures of a character or want two characters in one scene.

Limits

• Small accessories, collar shapes and fine patterns can change, especially with a strong style prompt.
• Upper-body references do not define lower-body clothing.
• Facial detail can suffer when the reference face is very bright.
• Text or signature-like marks in a reference can be copied into the output. Crop them out first.

Troubleshooting

The character looks nothing like the reference

You probably loaded the file with LoraLoader. Remove it and use the SnapRef node, which reads both LoRA branches correctly.

The SnapRef node is missing

Check the folder path is exactly ComfyUI/custom_nodes/snapref/ with the node files directly inside, not one folder deeper. Restart ComfyUI and check the console for import errors.

My style tags are ignored

Switch the mode to Character. In Character & Style mode the reference sets the style.

The second reference changes the outfit

Lower its weight (try 0.4) or set it to 0 to test without it. Make sure you are in separate layout, where weights apply.

The pose barely changes from the reference

Lower fidelity slightly and describe the pose clearly in the prompt.

A signature or text appears in my image

It was copied from the reference. Crop or clean the reference picture and run again.

FAQ

Is SnapRef free?

Free for non-commercial use. The weights use the CircleStone Labs Non-Commercial License v1.2. The code is Apache-2.0. SnapRef is not affiliated with CircleStone Labs.

Does it work with other models like SDXL or FLUX?

No. It is built for Anima base v1.0.

How many reference pictures can I use?

Up to eight in the simple node, all of the same character. The advanced node has three image inputs with weights.

Full file or FP8?

Full (382 MB) if you can spare the space. FP8 (195 MB) is fine for most use, with slightly different results.

Should I upgrade from v1a?

Yes if you want per-picture weights, Character mode or the two-character workflow. The authors rate its reference following better, while the metrics are about level.

How much VRAM does it need?

The page does not say. The adapter is small, so it should fit wherever Anima base already runs.

Sources

• SnapRef v2a on Hugging Face (files, nodes, workflows, settings, licence, limits)
• ComfyUI Wiki: SnapRef v2a
• ArcAnima/SnapRef (training and evaluation code)
• circlestone-labs/Anima