Noctaluna Squeezes Alibaba's Noct-Q-Anime Into 8GB GPUs for Anime Art
A community fine-tune of Alibaba's compact Qwen-Image-2.1 delivers uncensored anime illustration in an INT8 build that runs on 12GB consumer GPUs.
Takeaways
- Noctaluna released an uncensored anime fine-tune of Qwen-Image-2.1 with 6,000+ downloads.
- The INT8 build weighs 7.3 GB and runs on 8-12 GB VRAM cards in ComfyUI.
- Base model shrank from 20B/40.9GB to a 7B visual generator while adding native RGBA transparency.
- Recommended settings: 25 steps, euler simple, CFG 3, prompts starting with "An anime illustration of..."
- No trigger word needed; supports two-character scenes, explicit content, and long detailed prompts.
- Locked to Qwen Research License, non-commercial use only, unlike the Apache 2.0 predecessors.
Noctaluna has published Noct-Q-Anime, a community derivative of Alibaba’s Qwen-Image-2.1 built for anime illustration. The checkpoint stores the visual-generator weights in INT8, merges in anime-focused transformer weights, and removes the base model’s content restrictions. Its Hugging Face listing recorded more than 6,000 downloads within the first few days.
The author targets ComfyUI systems with 8 to 12GB of VRAM. The diffusion checkpoint occupies 7.3GB, although peak memory also depends on image resolution, batch size, model offloading, and the separate text encoder and VAE.
Seven billion parameters cut the footprint
The checkpoint builds on Qwen-Image-2.1 notes, Alibaba’s latest downloadable image model. Alibaba reduced the visual generator from roughly 20 billion parameters in the previous Qwen-Image release to 7 billion, cutting its BF16 weight size from 40.9GB to 14.2GB.
| Qwen-Image-2.1 component | Specification |
|---|---|
| Visual generator | 7 billion parameters |
| Architecture | 32 single-stream diffusion transformer layers |
| BF16 weights | 14.2GB |
| INT8 weights | 7.26GB |
| Supported tasks | Image generation, editing, and transparent output |
A diffusion transformer, or DiT, progressively converts noise into an image representation. BF16 stores each weight with 16 bits, while INT8 uses 8-bit integers to reduce storage and memory requirements. Quantization can affect output quality, but it makes the visual generator practical on a wider range of consumer hardware.
Alibaba reports an aggregate score of 60.28 in its published model comparison, ranking Qwen-Image-2.1 above the other downloadable-weight models included there. The release also supports images with transparency and arrived with integrations for ComfyUI and Diffusers.
The merge retunes style and removes restrictions
Noctaluna’s derivative preserves the INT8 format and merges transformer weights tuned for anime illustration. Its model card describes support for:
- Character portraits, full-body illustrations, and wide background compositions
- Long prompts containing specified poses, camera angles, clothing, and scene details
- Two-character compositions with distinct instructions for each subject
- Fictional adult characters in clothed, erotic, nude, or explicit scenes
Because the checkpoint has no trigger word, prompt wording controls the visual domain. The author recommends beginning with An anime illustration of... and continuing in full sentences. Omitting the word anime can shift the output toward photographic imagery.
| Setting | Recommended value |
|---|---|
| Steps | 25 |
| Sampler | euler |
| Scheduler | simple |
| CFG | 3 |
| Negative prompt | The prompt bundled with the workflow JSON |
Removing the base model’s content restrictions leaves moderation to the surrounding application. The model card states that its depicted characters are adults and that no real people are shown. Public or shared deployments would need their own input filters, output checks, access controls, and usage policies.
Three files make the local stack
Local inference requires the diffusion checkpoint, Qwen3-VL 8B text encoder, and Qwen-Image-2.1 VAE. With ComfyUI 0.37 or newer, the three files go into the standard core-node directories:
models/diffusion_models/NoctQA_V1_int8_convrot.safetensors
models/text_encoders/qwen3vl_8b_int8_convrot.safetensors
models/vae/qwen_image_2.1_vae_bf16.safetensors
The Qwen3-VL encoder converts long, multi-clause prompts into representations the image model can use. The VAE decodes the generated latent representation into final pixels. Runtime memory therefore exceeds the 7.3GB diffusion-checkpoint size, particularly at higher resolutions or without CPU offloading.
Other precision options, including FP8, FP16, BF16, and INT4 builds, are hosted on the author’s Civitai page rather than Hugging Face. Lower-bit variants reduce memory use, while higher-precision builds require more capable hardware.
Commercial use needs a separate agreement
Alibaba released the original Qwen-Image in August 2025 and the later Qwen-Image-2512 under Apache 2.0, which generally permits commercial use when its conditions are followed. Qwen-Image-2.1 instead uses the Qwen Research License.
The Qwen-Image-2.1 announcement describes the release as open source, while its license prohibits use for any commercial purpose without a separate agreement. That restriction is substantially narrower than Apache 2.0 and other permissive open-source licenses.
Noct-Q-Anime inherits the same terms as a derivative work. Personal and non-commercial research use is permitted under the model card, while paid products, commercial services, and other business uses require an agreement with Alibaba.
From 20B to 7B, specialization gets cheaper
Cutting the visual generator from roughly 20 billion parameters to 7 billion lowers the storage, memory, and compute needed to create and run specialized derivatives. The earlier model’s 40.9GB BF16 transformer limited comfortable local use, while an INT8 Qwen-Image-2.1 checkpoint fits workflows already used for SDXL and quantized Flux models.
Anime pipelines, character-sheet generators, and storyboard tools can now use a Qwen-based checkpoint on more consumer systems. The practical limits remain clear: the full stack consumes more memory than the diffusion file alone, removed safeguards require application-level moderation, and commercial deployment remains subject to Alibaba’s research license.