WooMix (Anima)

作者：Woozzii

Four checkpoints, one lineage, blended until the hands worked. v1.80 is a merge anchored on WooMix v1.5 and pulled toward more detail, more contrast and better anatomy. Same architecture, same text encoder, same VAE, same settings as v1.5 — different weights, and that is the whole point. This is the third release in the line. v1.75 proved the direction and drifted on consistency. v1.80 fixes the drift: the donor that was fighting the mix is out, the lead donor carries more weight, and the blend strength came out of a sweep instead of a guess. ✦ What you get

• Hands that survive a crop — fingers hold their count and their shape at full resolution, not just in a thumbnail

• Two characters that stay two characters — separate bodies, separate faces, no feature-borrowing when the frame gets busy

• Shadows that stay open — dramatic lighting keeps detail instead of collapsing to black

• Skin, fabric and hair with texture — grain and sheen that are still there when you zoom in

• One 5.2 GB full checkpoint — UNET, text encoder and VAE already inside. Nothing else to download

✦ What it measures against v1.5 Scored on a fixed 52-image paired benchmark: six prompt categories — hands, two characters, action, material texture, busy background, expression — at three seeds each, plus eight seeds of one character prompt to test whether the same girl comes back.

• Identity, worst seed — 0.898 on v1.5, 0.905 here

• Identity, 10th percentile — 0.910 on v1.5, 0.921 here

• Prompt adherence (CLIP) — 32.3 on v1.5, 33.2 here

• Speed — 17.2 s per 1024² render on v1.5, 17.4 s here

Identity is DINOv2 cosine similarity on an upper-centre crop — a consistency proxy, not a face-recognition verdict. Adherence is CLIP logits — an alignment proxy, not a taste score. Neither replaces your eyes; both say this build does not cost you consistency to gain detail. Honest trade: a 0.30 blend of the same donors is measurably more consistent — 0.919 worst-seed, 0.937 at P10. If rock-steady same-face output matters more to you than the extra detail, say so and that variant gets published too. v1.80 sits at 0.37 because that is where the detail arrives. It is a choice, not a free lunch. ✦ Recommended settings

• CFG 4.0–5.0 — start at 4.5

• Steps 30–40 — 30 for drafts, 35 for finals

• Sampler — er_sde for clean anime lines; DPM++ 2M SDE for more variation between seeds

• Scheduler — normal, or beta if you want it softer

• VAE — the bundled qwen_image_vae, always. Swapping it is the fastest way to broken colour

• Resolution — 1024×1024 and up

✦ Negatives, kept short Anima reads negatives through a Qwen text encoder, not CLIP. Long walls of negative text and bracket weights confuse it and flatten detail. Lowercase, spaces not underscores, no weight brackets. Minimum worst quality, low quality, lowres Standard worst quality, low quality, lowres, bad hands, missing fingers, watermark, signature, monochrome Multi-character and hands worst quality, low quality, lowres, bad anatomy, extra fingers, fused fingers, extra limbs, deformed, watermark, signature, text, 2koma, 4koma More is not better. Start at the minimum and only add when you are actually fighting an artefact — a long negative drains colour too. ✦ Merge details and licence Four Anima checkpoints, task arithmetic, blend strength 0.37, anchored on v1.5 so the output keeps its lineage instead of drifting into someone else's style. The recipe and the donor list are available on request — permissions on other people's checkpoints are theirs to give, not mine to publish. Built on Anima by CircleStone Labs. Released under the CircleStone Labs Non-Commercial License v1.2: no commercial use, no paywall, no reselling. ✦ Changelog

• v1.80 — blend 0.37, weaker donor removed, lead donor weighted up. Higher detail, measured consistency held.

• v1.75 — first four-way merge at 0.30. Proved the direction, drifted on character consistency.

• v1.5 — the base this whole line is built on.