CLIP is back on Anima, because CLIP is eternal.
A new ComfyUI node adds CLIP guidance to Anima, cutting 'horny' bias by 50% and improving prompt adherence.
A developer known as Anzhc has released a new ComfyUI node that brings CLIP-based 'Modulation Guidance' to the popular Anima anime-style image generation model. This technique, originally developed by researchers at Yandex and Adobe, integrates a separate CLIP text encoder (like CLIP-L) to guide the diffusion denoising process, acting as an additional steering mechanism beyond the model's native text understanding. The community-driven implementation addresses several well-known quirks of the Anima model, offering users a tool to gain more precise control over outputs.
Initial testing shows the Modulation Guidance node delivers tangible improvements: it reduces Anima's strong default biases, such as an unwanted tendency to generate beach/ocean backgrounds or overly sexualized figures ('less unprompted horny'), even when not prompted. It also improves compositional stability and reduces color bleeding between elements. While it doesn't support advanced features like SDXL's prompt weighting, the node offers a straightforward way to make Anima generations more reliable and adherent to user intent, marking a significant step in community-driven model refinement for open-source AI art tools.
- Integrates Yandex/Adobe's 'Modulation Guidance' technique via a new ComfyUI node for the Anima model.
- Uses external CLIP text encoders (e.g., CLIP-L) to steer generation, reducing model-specific biases and color leaks.
- Reportedly cuts down on Anima's default 'beach scene' and '1girl' biases, improving prompt adherence and composition.
Why It Matters
Gives AI artists more control over finicky open-source models, reducing unwanted biases and improving workflow reliability.