Text to image models have made significant progress in image quality, prompt understanding, and text rendering. However, generating graphics with dense text and intricate layouts remains challenging. An image can look visually appealing while still containing unreadable text, missing information, or poorly organized elements.
 |
| MOSS Image model Showcase (ref- Official page) |
MOSS Image,(Base and SFT variants ) released by the OpenMoss Team, make complex text-heavy graphics easier to generate by combining a specialized image model with an agent that improves prompts and plans layouts. Its generation model is a 5B-parameter single-stream Diffusion Transformer (DiT), pretrained from scratch for dense bilingual text and multi-element graphics.
Instead of relying entirely on a natural-language prompt to determine where everything should go, MOSS-Image can work with structured layout specifications. Its prompt enhancement module can also expand brief requests into more detailed instructions, helping the model produce graphics with clearer text placement, content hierarchy, and spatial organization. More detailed information can be found be accessing their research paper.
 |
| MOSS Image model architecture(ref-Official paper) |
There are two variants released-
(a) MOSS-Image-Base - It is the foundational text-to-image model in the MOSS-Image family. It supports three ways to describe the desired output- Natural-language prompts that describe the graphic in ordinary language. Tag sequences that specify relevant elements and attributes through structured tags. JSON layouts that define text placement and object arrangement using structured layout instructions. This makes Base suitable for workflows where users want flexibility in how they specify the composition.
(b) MOSS-Image-SFT - It is the supervised fine-tuned variant derived from MOSS-Image-Base. It uses the same architecture and inference pipeline while retaining support for natural-language prompts, tag sequences, and JSON layouts. The key distinction is that SFT has undergone additional supervised fine-tuning, whereas Base is the foundational model.
Installation
1. First you need to have the ComfyUI installation. If already have, just update it from the Manager tab.
2. Download Moss Image Base and Moss Image SFT from hugging face repository. Save it into ComfyUI/models/diffusion_models folder.
-moss_image_base_bf16.safetensors
-moss_image_base_int8_convrot.safetensors
-moss_image_sft_bf16.safetensors
-moss_image_sft_int8_convrot.safetensors
3. Download text encoders models. Save all of them into ComfyUI/models/text_encoders folder.
-qwen3.5_9b_bf16.safetensors
-qwen3.5_9b_int8_convrot.safetensors
-qwen3.5_9b_w4a8_gptq.safetensors
4. Download vae (flux2-vae_small.safetensors) model. Its the tiny variant of flux2-vae.safetensors. Save it inside ComfyUI/models/vae folder.
5. Download color depth lora (moss_image_color_depth_control.safetensors ). Its a control adapter that do color region mapping with depth map. Save it into ComfyUI/models/loras folder.
6. Restart and refresh ComfyUI.
Workflow
1. Download the Moss Image (Base/SFT) workflows from our Hugging face repository.
-moss_image_base_t2i.json (Base variant text to Image)
-moss_image_base_t2i_bboxes.json (Base variant text to Image with Bbox added feature)
-moss_image_t2i.json (SFT variant text to Image)
-moss_image_t2i_bboxes.json (SFT variant text to Image with Bbox added feature)
2. Drag and drop into ComfyUI.
3. Load the diffusion models, vae, text encoders into their relevant nodes.
4. Put the positive prompts into prompt box.
5. Configure the KSampler settings-
Moss image base settings(use this for finetune)-
Sampler- euler
Steps- 50
CFG- 3
shift-3
Neg prompt- no
Moss image SFT settings(for faster image generation)-
Sampler- euler
Steps- 50
CFG- 3
shift-5
Neg prompt- no
6. Hit the run to start generation.
