🎨 ANIMA / Qwen Encoder — Prompt Rules & Best Practices
ANIMA Base v1.0 · ComfyUI · Tested with euler_ancestral_cfg_pp + karras · CFG range 1.0 ~ 5.0
ANIMA uses a Qwen LLM-based text encoder instead of CLIP. This is a fundamental architectural difference that changes how prompts are processed.
CLIP boosts attention via weighted syntax — Qwen reads your prompt more like natural language, literally. Many SD / NoobAI prompting habits actively break ANIMA outputs. These are the rules that matter.
🚫 What To Never Do
① Weighted Syntax — Absolute Ban
❌(dramatically billowing cloth:1.2)
❌(all fabric blown in wind direction:1.3)
✅dramatically billowing cloth
Qwen reads (tag:1.3) as literal text → embedding collision → detail collapse. Just write it plainly.
② Comment Separators — Token Waste
❌// --- Hair --- / # Background Section
✅BREAK — the only separator that actually works
BREAK is natively understood by the sampler pipeline and properly separates attention regions.
③ Repeating the Same Concept 3+ Times
❌Describing wind across 6 separate lines → attention floods that region → all other detail collapses
✅strong wind, all fabric and hair blown rightward, ribbons and streamers mid-air
Say it once, say it well. Qwen gets it.
④ Spatially Conflicting Tags
❌foreshortened staff + staff larger than character
❌extreme close-up + full body visible
✅Pick one spatial directive. Use it. Commit.
Conflicting tags can't be resolved — the model compromises on texture, causing detail mush.
⑤ Blur-Inducing Tags
❌fabric edges motion blur / motion blur / speed lines on fabric
✅For wind/motion: blown, swaying, mid-air, trailing, swept
Motion blur tags don't stay local — they bleed into the entire render.
⑥ Token Limit — Stay Under 300, Aim for 200
⚠️Over 300 tokens → rear tags silently dropped + embedding averaging → everything looks generic
✅Under 200 tokens → clean, sharp, faithful output
If your prompt feels long, it is too long. Cut it.
⚙️ CFG — It's Inverted (Biggest Surprise)
4.0 ~ 9.0
❌ Noise explosion
color saturation collapse
painterly mush
← standard SD range
1.0 ~ 2.0
✅ Clean lines
stable detail
faithful to prompt
← ANIMA sweet spot
Qwen embeddings are already semantically dense and powerful. Amplifying them with high CFG pushes the sampling path into instability. Low CFG lets the model's natural guidance do its job.
Coming from SD where CFG 7 is standard? On ANIMA, CFG 1.6 is your new normal.
✅ Recommended Base Settings
Sampler euler_ancestral_cfg_pp / dpmpp_2m_sde_gpu
Scheduler karras / beta
Steps 25 ~ 30
CFG 1.0 ~ 2.0
Separator BREAK (between major prompt sections)
Token count under 200 recommended / hard limit ~300
📋 Quick Reference Cheat Sheet
🚫 Never Do This
❌ (tag:weight) syntax
❌ // --- comments ---
❌ Same concept 3+ times
❌ Spatially conflicting tags
❌ motion blur tags
❌ CFG above 3.0
⚠️ Prompt over 300 tokens
✅ Always Do This
✅ Plain natural language tags
✅ BREAK for section separation
✅ One concept, said once clearly
✅ One spatial directive at a time
✅ blown / swaying / mid-air
✅ CFG 1.0 ~ 2.0
✅ karras scheduler
ANIMA is not a drop-in replacement for NoobAI or Illustrious. It rewards clean, concise, conflict-free prompts and punishes the SD habits we've all built up over years.
Once you internalize these rules, the model's output quality jumps significantly — the architecture is genuinely capable. It just needs to be spoken to differently.