Qwen-Image 2.1 Omni Fits Into 4 GB of VRAM With Almost No Speed Loss

If you have ever deployed image generation locally, you know the usual pain point: one model is needed to create images, another to edit them, and together they do not fit into the VRAM of a typical graphics card. ChatR Labs is trying to solve both problems at once. It has released Omni, a LoRA adapter for Qwen-Image-2.1, along with a set of quantized builds, the lightest of which uses about 4 GB.

What Is Inside

Omni is a single rank-128 adapter spanning 224 modules, and the adapter itself weighs 671 MB. It teaches the base model two tasks at once: generating images from scratch based on text and editing an existing image from an instruction. There is no need to swap weights between modes. It retains everything the base model is valued for: a 7B DiT, native 2K resolution, RGBA transparency, and improved typography.

The playground accepts up to three reference images, although the underlying Qwen-Image-2.1 model supports ten. If you need a multi-reference build, you will have to run it locally and test it yourself. The "Uncensored" label should also be read carefully: the filters have been relaxed, so responsibility for its use lies with you and your product.

Memory and Time Costs

The numbers are the most interesting part. The post's author tested the model on an RTX 5090 at 768x1152 and 35 steps:

• BF16: about 14 GB of VRAM, roughly 28 seconds
• Q5_K_M: about 8 GB, the same 28 seconds
• NVFP4: about 4 GB, about 30 seconds

In other words, quantization barely affects speed. The trade-off is image quality, and the exact impact depends on the scenes you generate. The build lineup is extensive: Q4_0 at 4.20 GB, Q4_K_M at 4.47 GB, Q5_K_M at 5.21 GB, Q6_K at 6.00 GB, Q8_0 at 7.69 GB, FP8 at 7.12 GB, and BF16 at 14.23 GB. Separate FP8 and NVFP4 versions are available for ComfyUI.

A caveat is necessary: the measurements were taken on a top-end 5090 and reported by a single author. NVFP4 is hardware-accelerated on the Blackwell architecture, so timing will differ on older cards. Do not treat the table as a performance guarantee for your own hardware.

Settings and Pitfalls

The recommended settings are 35 steps, true CFG 5.0, the er_sde sampler with the simple scheduler, 1024x1024 resolution, and KV cache enabled. There is a non-obvious detail involving the negative prompt: it must contain a single space. If it is left empty, true CFG simply turns off and the result becomes noticeably worse, without producing any error.

The second trap concerns the files. The ComfyUI and Diffusers builds are not interchangeable: they contain 265 and 297 tensors, respectively. Use the wrong format and the model will not load. You only need to download the transformer; the text encoder and VAE come from the base Qwen-Image-2.1 model.

What to Do With It

If you have a graphics card with 6–8 GB of VRAM, this is an opportunity to try local image generation and editing without the cloud or per-request fees. For a product, it provides a way to prototype image editing inside your own environment, where data does not leave your infrastructure. Start by testing quality on your typical tasks in the playground: it works without an account, although the queue runs on two 4090s and may be long.

Then choose a quantization level that fits your memory. The adapter is currently at version v0.1, so for production it is more of an experimental foundation than a finished solution.

And be sure to review the base model's license and the legal risks of an "uncensored" model before exposing it to users.