Anima Mod Guidance (model patch)

Anima's default output is famously plain. It's not a flaw so much as a design choice - the base model was never aesthetic-tuned - but it means new users generate one bare-prompt image, conclude the model looks bad, and bounce. The community's own fix is mostly about prompting: find artist tags, use a latent upscale pass. Anima Mod Guidance attacks the same problem from a different angle - instead of steering through the prompt, it steers through the model's own internals.

How it steers, without touching your prompt

This is a MODEL → MODEL patcher, not a sampler. It wires a learned adapter - a small pooled_text_proj MLP - into Anima's AdaLN timestep embedding. You give it quality_tags (what you want more of) and quality_neg (the counter-pole), it encodes both through CLIP, and the adapter projects the difference between them into a guidance delta that gets injected per transformer block according to mod_w_profile. It's a second, independent lever alongside your actual prompt and CFG - closer to a quality dial baked into the architecture than to more prompt text.

Inputs and outputs that matter

• model - the model to patch.
• clip - required, to encode quality_tags/quality_neg.
• quality_tags (default highres, best quality, score_7) and quality_neg (default score_1, score_2, score_3, worst quality, lowres, old, bad hands, bad anatomy) - the steering axis. Anima genuinely was trained with Pony-style score_x tags, so score_7 in the default is real, not decorative - just don't push it to score_9, which tends to pull the output toward furry/e621 territory since that's where the score classifier's training data leans.
• mod_w_profile - the per-block schedule, and the one setting worth understanding before you touch anything else:
  • off disables guidance entirely - no download, no hook, zero overhead.
  • step_i8_skip27 (default) protects the first eight blocks and the final compensation block from steering, applying the guidance to everything in between. Best overall quality, but the tooltip is honest that it can show minor anatomy drift on LoRAs that are already prone to it.
  • step_i14 is the safe fallback - steers only from block 14 onward, more conservative, less expressive.
  • uniform_w3 recovers old pre-update behavior. Not recommended - the tooltip flags it as prone to pink-collapse.
• positive / negative - CONDITIONING inputs. This node only reads them to compute the steering delta; wire the same conditioning you're feeding your actual sampler here too.

Output is a patched MODEL, ready to feed into any sampler - the base or Advanced Spectrum samplers, the SPEED node, or a stock KSampler.

Installing it

ComfyUI Manager - search SpectrumKSampler, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/sorryhyun/ComfyUI-Spectrum-KSampler

Restart ComfyUI. On first actual use, the ~12MB pooled_text_proj adapter auto-downloads to ComfyUI/models/anima_mod_guidance/ - expect a short one-time pause the first time you run a graph with this node in it.

Common issues & troubleshooting

Seeing anatomy drift on a specific LoRA. That's the known tradeoff of the default step_i8_skip27 profile on drift-prone LoRAs. Switch to step_i14 - it steers a narrower band of blocks and stays closer to the trained manifold.

Output looks washed out toward a generic "polished anime" look. Overuse of score_x-style quality tags does exactly that - it narrows the output distribution toward whatever aesthetic the training data's classifier favored. If you're already specifying an artist tag or an unusual style, it's worth dropping the quality-tag prefix rather than fighting it with mod guidance.

Wondering if this works on a non-Anima checkpoint. The whole pack - mod guidance included - was built and tuned specifically for Anima's DiT. It may run on other DiT models that expose compatible sampler internals, but the pack doesn't claim quality or stability there.

Not sure it's doing anything at all. Unlike the integrated samplers, this node's clip input is required, not optional - so if it's wired at all, guidance is active. If you want to A/B it, set mod_w_profile = off rather than disconnecting clip.