<h1 id="anima-role-ref-kv-v2.0">Anima Role Ref KV · v2.0</h1>Generate new poses, outfits, expressions, and scenes from a single character reference image and a text prompt on **Anima Base v1.0**, without per-character training or trigger words.Required downloads: <code>anima-role-ref-kv-v2.0.safetensors</code> AND <code>anima-role-ref-kv-comfyui-setup-v2.0.zip</code> (Training Data / Attachments).v2.0 ships **BF16 weights** (50% smaller). If upgrading from v1.0, you MUST replace <code>ComfyUI/custom_nodes/anima-role-ref-kv</code> from the new setup ZIP and restart ComfyUI — older FP32-only nodes will reject BF16 files. The updated node supports both v2.0 and v1.0. v2.0 is 15% faster by using BF16 for all calculations.[<a target="_blank" rel="ugc" href="https://huggingface.co/AnticPan/Anima-Role-Ref-KV">Hugging Face Repository (Diffusers + ComfyUI)</a>] · [<a target="_blank" rel="ugc" href="https://huggingface.co/spaces/AnticPan/Anima-Role-Ref-KV">Online Demo Space</a>]<h3 id="what-changed-in-v2.0">What changed in v2.0</h3>- **Higher character resemblance**: normalized head similarity <code>1.23 -&gt; 1.72</code> across 3,008 development evaluations.- **Far fewer washed-out outputs on natural-language prompts**: unintended low-color rate drops from <code>31.1%</code> to <code>3.2%</code>.- **Tradeoff**: overall tag hit rate is <code>90.3%</code> (vs <code>91.2%</code> in v1.0); stronger preservation of reference rendering and garment details can sometimes resist heavy outfit or style changes.<h3 id="quick-start-(comfyui)">Quick Start (ComfyUI)</h3>1. Extract <code>anima-role-ref-kv-comfyui-setup-v2.0.zip</code> and copy <code>custom_nodes/anima-role-ref-kv</code> into <code>ComfyUI/custom_nodes/</code>.2. Put <code>anima-role-ref-kv-v2.0.safetensors</code> into <code>ComfyUI/models/loras/</code>.3. Restart ComfyUI and load <code>workflows/anima-role-ref-kv-v2.0.json</code> (requires <code>anima-base-v1.0.safetensors</code>, <code>qwen_3_06b_base.safetensors</code>, <code>qwen_image_vae.safetensors</code>).### Recommended Settings &amp; Prompting- **Sampling**: **28 steps, CFG 4, BF16**, ~1 MP <code>832 x 1216</code>, <code>1024 x 1024</code>, or <code>1216 x 832</code>, multiples of 16).- **Tag prompt**: <code>highres, safe, 1girl, solo, upper body, sitting, bench, park, tree</code> (do not repeat the reference character's existing hair/eye traits).- **Natural-language prompt**: <code>safe, 1girl. Upper body shot from the waist up. She is sitting on a bench in a park with trees.</code>- **Negative prompt**: <code>worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration</code>Independent modified derivative of Anima, subject to the [CircleStone Non-Commercial License](<a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima-Base-v1.0-Diffusers/blob/main/LICENSE.md">https://huggingface.co/circlestone-labs/Anima-Base-v1.0-Diffusers/blob/main/LICENSE.md</a>); not an official CircleStone Labs product.

Anima-Role-Ref-KV

v2.0 (BF16 weights, 50% smaller). Higher character resemblance (1.23 -&gt; 1.72) and far fewer unintended washed-out outputs on natural-language prompts (31.1% -&gt; 3.2%). IMPORTANT: Replace custom_nodes/anima-role-ref-kv from <a target="_blank" rel="ugc" href="http://anima-role-ref-kv-comfyui-setup-v2.0.zip">anima-role-ref-kv-comfyui-setup-v2.0.zip</a> and restart ComfyUI before loading v2.0.

Files with SHA256: e9409a608bb33e23106f0143b995930fd035efb976d519c8cb9429e58f196bb4