Use directly: load it as the diffusion model — no LoRA or patch nodes needed.

File: anima29b-p1-turbo1.1.safetensors
Size: ~5.4 GiB — full model (merged UNET, 40 blocks, BF16)

Goal: turbo speed (8 steps) with Anima 2.9B Preview 1-tier quality, character/artist knowledge and NSFW performance.

How it was merged

Shared blocks 0.8 × Turbo 1.1 + 0.2 × 2.9B P1; the 12 extension blocks unique to 2.9B P1 are kept as-is. The 0.8 strength was settled by full-scene same-seed comparisons: SFW and 100%-crop details on par across ratios, 0.8 slightly better in NSFW anatomy; no spontaneous mosaic / black bars at any ratio.

Workflow

UNETLoader (this model) + CLIPLoader (qwen_3_06b_base.safetensors) + VAELoader (qwen_image_vae.safetensors) → two CLIPTextEncode (positive/negative) → KSampler → VAEDecode → SaveImage. No LoRA needed.

Recommended parameters

steps 8 (8-12)
sampler euler + sgm_uniform
cfg: 1.0 without negative; 1.2 with negative (range 1.0-1.6 — negatives take effect as soon as cfg > 1)
resolution: 1024×1024 baseline; 1264×848 landscape / 848×1264 portrait tested fine (~14-17 s/image)

Prompting (quality tags work)

Same as the turbo family — masterpiece, best quality etc. work:

masterpiece, best quality, <safe/sensitive/nsfw/explicit>, <count>, <character (series)>, @<artist>, <appearance tags>, <natural language sentences>

Character name right after the series tag; avoid underscores (hatsune miku, not hatsune_miku)
@artist strongly steers the style — recommended
Write 1-2 full natural-language sentences; for multiple characters describe each one's appearance next to them
Standing text-type negative text, letters, sign, logo, watermark, signature, username, artist name + cfg 1.2 recommended: noticeably reduces corner fake signatures / watermarks / logos (useless against text that belongs to the scene itself); for scenes without an artist tag, an empty negative + cfg 1.0 is fine
Not good at rendering readable text; for NSFW just write the rating and content tags directly

Sample images

All samples (s1/s3/s4/s5) are straight outputs of this model at 8 steps, cfg 1.2 with the text-type negative above. Full ComfyUI workflow metadata is embedded in the PNGs — drop into ComfyUI to reproduce.

2.9B-P1-Turbo1.1 (Unofficial Merge) - p1-turbo1.1