OpenMOSS releases 5B text-to-image model MOSS-Image with public weights and code, and ComfyUI supports it natively on day one
Synopsis
OpenMOSS has released MOSS-Image, a 5B text-to-image model, together with code and two weight versions, Base and SFT. Its stated strengths are layout control, text rendering inside images and prompts up to 4K tokens, and it reports a peak of about 3.65 GiB of allocated PyTorch GPU memory for one 1024×1024 image at 50 steps. ComfyUI, SGLang and vLLM-Omni supported it on release day, and community projects followed with LoRA training, distillation and a C++ inference engine.
Interpretation
Open-source text-to-image work gains a 5B-scale option: MOSS-Image ships two weight versions, Base and SFT, alongside its code, ready to download. The release notes list the 5B backbone, public weights and code, and easy fine-tuning in one set of highlights, and give entry points for ComfyUI, SGLang and vLLM-Omni at the same time. ComfyUI, SGLang and vLLM-Omni supported it on release day, and community projects followed with LoRA training, 8-step distillation and a C++ inference engine.
The model takes layout as structured input: a JSON description of the canvas, element captions and bounding boxes fixes where text and objects appear. An ordinary text-to-image prompt leaves placement to the model, whereas here each element's label, caption, text content and bounding box can be written into the prompt. Bounding boxes use pixel coordinates centered on the canvas in yxyx format, running from −512 to +512 per axis on a 1024×1024 canvas, and must match the requested output size.
It also runs on memory-constrained machines: low-memory mode keeps DiT weights in CPU RAM and loads text-encoder and VAE weights from disk as needed. Keeping the whole model resident on the GPU fails when memory runs short, and staged loading lowers that barrier; sequential fallback mode trades slower inference for lower usage. For one 1024×1024 image at 50 steps, measured peaks were 3.65 GiB allocated and 3.85 GiB reserved PyTorch GPU memory, with 14.00 GiB of task system RAM including file cache.
Perspective
The demonstrated scope is two public weight versions, JSON layout input, prompts up to 4K tokens, and measured peaks for one 1024×1024 image at 50 steps in low-memory mode; the intended readers are developers and researchers willing to deploy or fine-tune an image model themselves.
The material gives no comparison of image quality or text-rendering accuracy against comparable models; sequential fallback mode trades slower inference for lower usage, and compiled DiT is unavailable in both sequential and low-memory modes.
