OpenMOSS Releases MOSS-Image With Native Layout Control and 4K Prompts
The Apache-licensed 5B diffusion transformer pairs bilingual text rendering with structured JSON bounding boxes for precise composition.
In short
- OpenMOSS released MOSS-Image-Base and MOSS-Image-SFT, a 5B diffusion transformer family distributed under the Apache 2.0 license.
- The model natively interprets JSON layout prompts using bounding boxes to place specific objects and rendered text across canvases up to 2K resolution.
- Built with a Qwen3.5-9B text encoder in the FLUX.2 latent space, MOSS-Image accepts prompt inputs up to 4,000 tokens long.
- Day-0 ecosystem integrations include native ComfyUI workflow nodes, vLLM-Omni support, and LoRA training on Civitai.
The OpenMOSS team introduced MOSS-Image on October 16, 2026, launching a family of open-weights text-to-image models designed around a 5-billion parameter Diffusion Transformer (DiT) architecture. Available in both a foundational base version and a supervised fine-tuned checkpoint, the release pairs bilingual Chinese and English text generation with a structured JSON layout mechanism that lets creators position visual elements and typography across the canvas using coordinate bounding boxes.
Distributed under the Apache 2.0 license on Hugging Face and ModelScope, the release addresses persistent friction points in generative image pipelines: rendering legibly grouped text, respecting spatial composition instructions, and maintaining adherence to descriptions stretching up to 4,000 tokens.
Architecture and Native Layout Control
At the technical core of the system is a 5B single-stream DiT operating within the FLUX.2 latent space. According to integration notes analyzed in ComfyUI pull request #17095, the model employs a stock Qwen3.5-9B text encoder and pairs it with the FLUX.2 variational autoencoder (VAE) decoder. The bundled VAE operates in float32 precision, while the core transformer and text encoder run in bfloat16.
Unlike traditional diffusion workflows that rely exclusively on natural-language descriptive paragraphs or tag lists, the MOSS-Image project documentation outlines native layout-to-image generation. Users can pass structured JSON strings containing a global caption alongside an array of bounded items. Each item includes bounding box coordinates, a label, and any specific text string to be rendered directly into the illustration.
The layout system utilizes centralized absolute coordinates (bbox_coord_space: "abs_centralized") in a [y_min, x_min, y_max, x_max] format relative to the canvas midpoint. On a standard 1024×1024 pixel canvas, each axis ranges from −512 to +512 pixels. This mechanism allows developers to programmatically arrange signs, poster titles, and focal objects before generation begins, rather than hoping a stochastic diffusion process places them appropriately.
To bridge the gap between creative briefs and spatial syntax, the developers released the MOSS-Image Prompt Enhancer, an agent skill aimed at GPT-6-class models. The bundled tool translates pixel-based visual drafts into the integer-centered JSON schema understood by the diffusion pipeline.
Base vs. SFT Checkpoints
The project ships in two distinct variants: MOSS-Image-Base and MOSS-Image-SFT. Both models share identical architectures and sampling settings, but serve different phases of deployment:
- MOSS-Image-Base: The raw pretrained base model, trained across heterogeneous GPU clusters. OpenMOSS recommends Base as the starting foundation for downstream domain adaptation and Low-Rank Adaptation (LoRA) fine-tuning.
- MOSS-Image-SFT: Supervised fine-tuning performed on curated datasets. The official repository documentation demonstrates that SFT produces clearer text grouping and a more organized whiteboard layout compared to the unaligned base model.
For inference, both models rely on a Flow-matching Euler sampler configured by default for 50 steps, a classifier-free guidance (CFG) scale of 3.0, a shift value of 5, and negative-branch layer skipping at index 22. Community platform Civitai noted that its web generator supports MOSS-Image at 25 steps with CFG 3, while hosting full LoRA training support in its cloud Training Studio on both Base and SFT starting points.
Ecosystem Adoption: ComfyUI, vLLM, and Quantization
Adoption across open-source deployment engines occurred within hours of release. Comfy-Org merged native pipeline support directly into ComfyUI core on October 16, adding dedicated nodes including the MOSS-Image Layout Prompt node. This node connects bounding-box tools on the ComfyUI canvas to the model's spatial layout inputs.
To improve local execution on consumer hardware, Comfy-Org released repackaged weight formats, including an INT8 ConvRot diffusion model checkpoint measuring approximately 5.4 GB, alongside a Qwen3.5-9B text encoder available in 9.8 GB INT8 ConvRot or 7.3 GB W4A8 GPTQ. In a follow-up commit documented by ArtRealmAI, ComfyUI integrated fused SiTU input activation via comfy-kitchen 0.2.38, delivering an estimated 30 percent speedup for INT8 quantization runs.
Simultaneously, inference framework vLLM announced Day-0 support for MOSS-Image through vLLM-Omni.
Memory Modes and Local Requirements
For local Python execution via Hugging Face Diffusers, MOSS-Image requires Python 3.10 or newer, PyTorch 2.9 or higher, and a CUDA-capable NVIDIA GPU. The official repository includes a Gradio interface equipped with three explicit memory strategies:
- GPU Resident: Keeps all 5B transformer parameters, encoders, and VAE resident in video memory for maximum throughput.
- Component: Dynamically shuttles individual model components between system RAM and GPU VRAM during generation phases.
- Sequential: An aggressive fallback mode that minimizes VRAM consumption at the expense of lower generation speeds, disabling DiT graph compilation.
By releasing the weights under Apache 2.0 alongside production tooling, OpenMOSS provides creators and enterprises with a modular text-to-image foundation that balances fine-grained spatial control against accessible hardware requirements.
Questions people ask
What license governs the MOSS-Image weights and code?
Both the model weights (MOSS-Image-Base and MOSS-Image-SFT) and the accompanying pipeline code are released under the Apache License, Version 2.0.
How does MOSS-Image control object layout on the canvas?
Users pass a structured JSON prompt containing summary descriptions and an items list. Each item specifies bounding box coordinates using centered pixel values [y_min, x_min, y_max, x_max], allowing exact placement of objects and rendered text.
Which checkpoint should be used for LoRA training?
The OpenMOSS team recommends MOSS-Image-Base as the preferred starting point for custom task fine-tuning and LoRA training, though LoRAs trained on Base are also largely compatible with MOSS-Image-SFT.
What third-party components are used in the MOSS-Image architecture?
MOSS-Image combines a custom 5B Diffusion Transformer with a Qwen3.5-9B text encoder, operating within the FLUX.2 latent space using the FLUX.2 small VAE decoder.
Sources
- OpenMOSS-Team/MOSS-Image-SFTHugging Face, Oct 11, 2026, official source
- OpenMOSS/MOSS-ImageGitHub, Oct 16, 2026, official source
- OpenMOSS-Team/MOSS-Image-BaseHugging Face, Oct 11, 2026, official source
- Image | Text-to-Image Generation with Native Layout Controlopenmoss.ai, Oct 16, 2026
- MOSS-Image goes open source with bilingual text rendering and experimental layout controle/acc, Oct 17, 2026
- MOSS Image: Text That Reads Right, Prompts Up to 4K TokensCivitai, Oct 17, 2026
- MOSS-Image Goes Native in ComfyUIArtRealmAI, Oct 17, 2026
How this story was made: the newsroom picked it up from Reddit and GitHub, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (33 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 19, 2026 at 00:47 UTC
