Anima-Role-Ref-KV

作者：anticpan

Anima Role Ref KV · v3.0 Required downloads: anima-role-ref-kv-v3.0.safetensors AND anima-role-ref-kv-comfyui-setup-v3.0.zip (Attachments). v3.0 ships BF16 weights. Replace ComfyUI/custom_nodes/anima-role-ref-kv with the node from the new setup ZIP and restart ComfyUI — older nodes cannot read the pose image. The new node also runs v2.0 and v1.0. Generate a character in a chosen pose on Anima Base v1.0: one character image for identity and appearance, one optional pose image for pose and placement, and a text prompt. No per-character training or trigger words. [Hugging Face Repository (Diffusers + ComfyUI)] · [Online Demo Space] [AI Pose Gallery] (original; Chinese version) What's new in v3.0 Pose image input. Use a line-art pose guide or another single-person illustration, framed with the output's aspect ratio. The line-art guides in the training data come from [AI Pose Gallery] (original; Chinese version). Pose followed (automated judge, development set): line-art guides 92.6–97.7%, another character's image 86.4%, real illustrations 71.9%. One training seed. Faster ComfyUI node. Uses ComfyUI's selected attention backend (xFormers, SageAttention, FlashAttention or PyTorch) and FP32 attention on GPUs without native BF16 (GTX 10xx, RTX 20xx). The main workflow is Apply node + KSampler, so sampler, scheduler and other LoRAs (e.g. Turbo) stay under your control. Tips Body shape. The pose image's body (proportions, build, chest) can carry over. Use a pose figure close to the character's body. Male characters. Line-art guides often use a female figure. Write 1boy, male focus, flat chest (or male chest) in the prompt. Bright or white backgrounds. Add white background, simple background to the negative prompt, and name the place, time of day and lighting in the prompt (this negative prompt can make images darker and more saturated). Prompt. Describe outfit, expression and background; coarse pose words sitting, on chair, kneeling) are fine when they agree with the pose image. Character-only generation: use v2.0. v3.0 may follow editing requests (outfit, expression, style changes) less well than v2.0. NSFW images and NSFW poses were removed from the v3.0 training set; the model may still generalize to such content to some extent.

Recommended settings

• Sampling: 28 steps, CFG 4, BF16, ~1 MP 832 x 1216, 1024 x 1024, 1216 x 832, multiples of 16).
• Tag prompt: highres, safe, 1girl, solo, <outfit>, smile, outdoors, park, bench, tree, full body, sitting, on bench
• Negative prompt: worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration (add , white background, simple background if backgrounds turn white)

Independent modified derivative of Anima, subject to the CircleStone Non-Commercial License; not an official CircleStone Labs product.