MOSS Image Studio explains how to get exact text from the open MOSS-Image model

What the studio learned

MOSS-Image is a 5-billion-parameter image model released by OpenMOSS under the Apache 2.0 licence, and according to a dev.to post by the team behind MOSS Image Studio, it was trained with legible lettering as an explicit objective rather than an incidental one. The team has built a browser studio around the model and generated a few hundred poster images while preparing sample material; their write-up describes which variables actually decide whether the output is printable.

Their central point is that the same weights driven by a loosely written prompt still produce the usual smeared lettering. Four practices close the gap, listed in order of importance.

Four rules for reliable lettering

First, write the copy out word for word inside quotation marks and label each element's role, such as headline, line or footer. Those labels need to stay plain: the model treats ordinary language as image content and will sometimes print a role word into the picture.

Second, scale beats density. Five lines at readable size render far more dependably than a dozen lines of small type, and layouts that genuinely need dense text should be expected to take several attempts.

Third, placement must be enforced rather than hinted at. Asking for empty space in the upper third is advice the model can ignore, which is why the studio built a box editor: the user draws the text area on the canvas and the model fills it.

Fourth, everything in the prompt is potential picture. Naming a café can summon a café sign covered in invented lettering, and asking for "no signs" is itself only advice. The team's fix is to omit objects that tend to carry writing rather than request their absence.

The ComfyUI trap

For people running the model locally, the post names checkpoint confusion as the most common reason users conclude the model cannot handle text. ComfyUI's four text-to-image templates are built for the instruction-tuned SFT checkpoint, moss_image_sft_int8_convrot.safetensors, run at 25 steps with CFG 2.5 and the er_sde sampler. The separate Base checkpoint, moss_image_base_int8_convrot.safetensors, expects 50 steps, CFG 3 and the euler sampler instead. Pairing a template with the wrong checkpoint produces weak results that read as a model limitation.

The setup puts three files in three folders: the diffusion model under models/diffusion_models/; the text encoder qwen3.5_9b_w4a8_gptq.safetensors under models/text_encoders/, loaded through a CLIPLoader set to the moss_image type rather than qwen or flux; and the VAE flux2-vae_small.safetensors under models/vae/. Local ComfyUI suits anyone who wants the weights on their own GPU with no per-image cost, the team writes, but not someone who needs one poster tonight.

Known failure modes

The write-up is candid about limits. Paragraphs, legal lines and dense price tables make characters drift at small sizes and should be split into separate lines or dropped. Latin and Chinese scripts render well; rarer languages are a gamble and should be checked glyph by glyph before printing. Two generations from the same brief yield two different designs, so for a matching series the team recommends generating once and remixing in the box editor rather than re-rolling the prompt. A six-item price list will usually get each line's digits right but not in order of importance, which means reordering and rerunning.

The closing advice is procedural: write the copy in a text file before opening any tool, decide how many lines exist and how large each must be, generate a single image to see whether the model agrees with the plan, trim the text, and only then adjust the scene. Most failed posters, the team argues, are text problems disguised as rendering problems.

The post also notes that MOSS Image Studio is independent and not affiliated with OpenMOSS, and that its hosted studio gives new users three images without requiring a payment card.

Why it matters

Faithful lettering has long been one of the most stubborn weaknesses of image generators, which is why real-world design work still tends to end in a layout tool. An Apache 2.0 model that can render text well enough to print, combined with documented technique for prompting, placement and checkpoint configuration, pushes menus, price tags and posters toward something anyone can run on their own hardware. Equally valuable is the honest catalogue of failures, which sets realistic expectations, and the ComfyUI notes, which address a quieter problem: local misconfiguration producing verdicts of "can't do text" that are really setup errors.