Anima-turbo

Anima Turbo — anime generation + LLLite inpainting

A 4B anime-native turbo model — ≈6 s per 1024² image on an RTX 5090 at 10 steps, with a ControlNet-LLLite inpainting patch for masked edits.

What this repo is: the Anima Turbo denoiser, the LLLite inpainting patch, a Qwen3-0.6B text encoder and the Qwen-Image VAE — weights only, not a retrain. The settings below are the values these weights are actually run with day to day.

---

Samples

Anime is what this model is for — these are 10-step, guidance 1.0 renders.

<table> <tr><td><img src="https://huggingface.co/ChrisColeTech/Anima-turbo/resolve/main/samples/anime-rooftop.png" width="380" alt="anime girl on a rooftop at sunset"></td><td><img src="https://huggingface.co/ChrisColeTech/Anima-turbo/resolve/main/samples/anime-knight.png" width="380" alt="anime silver-armored knight in a ruined cathedral"></td></tr> <tr><td>_{**prompt:** `masterpiece, best quality, anime illustration of a girl in a school uniform standing on a rooftop at sunset, detailed cel shading, clean linework, dramatic sky` — 1024×1024, 10 steps, guidance 1.0, seed 21}</td><td>_{**prompt:** `masterpiece, best quality, anime illustration, a silver-armored knight with a glowing blue sword in a ruined cathedral, dramatic lighting, detailed` — 1024×1024, 10 steps, guidance 1.0, seed 7}</td></tr> <tr><td><img src="https://huggingface.co/ChrisColeTech/Anima-turbo/resolve/main/samples/anime-street.png" width="380" alt="anime background art of a japanese street in summer"></td><td><img src="https://huggingface.co/ChrisColeTech/Anima-turbo/resolve/main/samples/inpaint-source.png" width="380" alt="anime girl in a white dress in a flower meadow"></td></tr> <tr><td>_{**prompt:** `masterpiece, best quality, anime background art of a quiet japanese street in summer, blue sky, detailed clouds, vibrant colors, no people` — 1024×1024, 10 steps, guidance 1.0, seed 22}</td><td>_{**prompt:** `masterpiece, best quality, anime illustration, a young woman with long black hair in a white summer dress standing in a flower meadow, blue sky` — 1024×1024, 10 steps, guidance 1.0, seed 21}</td></tr> </table>

---

Inpainting (img2img + mask)

img2img on this model is masked inpainting through the LLLite patch, not strength-based restyling. Supply the source image, a white-on-black mask of the region to change, and a full prompt describing the desired result.

<table> <tr><td><img src="https://huggingface.co/ChrisColeTech/Anima-turbo/resolve/main/samples/inpaint-source.png" width="250" alt="source image"></td><td><img src="https://huggingface.co/ChrisColeTech/Anima-turbo/resolve/main/samples/inpaint-mask.png" width="250" alt="mask over head and torso"></td><td><img src="https://huggingface.co/ChrisColeTech/Anima-turbo/resolve/main/samples/inpaint-red.png" width="250" alt="inpainted result"></td></tr> <tr><td>_{**source** — the meadow render above, seed 21}</td><td>_{**mask** — ellipse over head and torso; white = repaint}</td><td>_{**result** — prompt changed to `…a young woman with vivid red hair…`, seed 43. Hair tone, eyes and bodice change; the meadow, pose, dress and framing are untouched.}</td></tr> </table>

What inpainting does and does not do here

• Recolouring and restyling inside the mask works well — hair colour, eye colour, garment detail. Everything outside the mask is preserved exactly.
• Adding new objects is unreliable. wearing a wide straw sun hat on the same mask produced no hat, only a slight expression change. LLLite is a control patch conditioning an existing composition, not an instruction editor — write the whole scene as you want it, and expect changes in the masked region to stay close to the underlying shapes.
• Without a mask, img2img is effectively a no-op — the control signal reproduces the input. Style-transfer prompts return the source image unchanged. If nothing is changing, check that a mask was actually sent.

Recommended settings

Values this build is run with in practice.

Supported modes: txt2img, img2img (masked inpainting)

Notes and gotchas

• strength here is ControlNet conditioning strength, not img2img denoise strength — it scales how hard the LLLite patch holds the source, and the engine also accepts it as control_strength. Lower it to let the masked region drift further from the original.
• Quality tags matter. This model responds strongly to booru-style prefixes (masterpiece, best quality, anime illustration, …); without them output drifts toward generic digital painting.
• Photographic prompts are off-domain. Asking for a studio photograph yields painterly, illustration-inflected results — use a photo model for that and this one for anime.
• ≈5.4 GB VRAM resident at 1024². Generation is fast enough (≈6 s) that iterating on prompts is interactive.

---

Layout

Components ship as separate files: the denoiser, the LLLite inpainting patch under split/model_patches/, the Qwen3-0.6B text encoder, and the VAE. Any loader that accepts explicit per-component paths and can apply a ControlNet- LLLite patch can consume this directly.

---

Files

---

Provenance

• Upstream base model: Anima Turbo v1.0 (anime-specialised, Qwen-Image family architecture)
• This build: redistributed with the inpainting patch, encoder and VAE alongside. Weights are not retrained here.
• License: left as unknown in this repo's metadata. Refer to the upstream model's license for redistribution and commercial-use terms.