Glyph-ByT5 Sampler

This is the sampler for the Glyph-SDXL pipeline - the node that actually generates the image, with your multilingual text laid into it. Glyph-ByT5 is a character-aware text encoder (fine-tuned ByT5, out of Microsoft research) fused with SDXL; the combination pushes design-image text accuracy from under 20% to near 90% on the paper's benchmark, and can render tens to hundreds of characters with automatic multi-line layout. This node is where that happens.

Fair warning before you wire it up: this is the least reliable corner of the pack. The author labels it "TEST FAILED ON MY PC," and the model comes with hard constraints. Go in expecting a research demo, not a workhorse.

How it works

The node takes a glyph_sdxl_model, the font_params from UL_Image_Generation_Glyph_ByT5_Font (your text lines, fonts and colors), a mask for where the text sits, and a set of background-description inputs. Glyph-ByT5's encoder handles the glyphs while SDXL renders the scene, and the node denoises the two together into a latent. The scheduler defaults to TCD at just 8 steps and CFG 1.5 - that's a distillation-style fast-sampling setup, so don't "fix" the low CFG; it's meant to run that way.

The inputs and outputs that matter

• glyph_sdxl_model (Glyph_SDXL_Model) and font_params (Glyph_SDXL_Font_Params) - the model and your text layout. Both required; font_params comes from the Font node.
• bg_prompt - the background scene description. Note the split: your text lives in font_params, your scene lives here.
• bg_class and bg_tags - a design category (posters, cards, invitations…) and descriptive tags that steer the background aesthetic. These reflect how Glyph-SDXL was trained on graphic-design data.
• mask (MASK) and mask_gap - where text goes and the spacing between regions.
• steps (8), cfg (1.5), scheduler (TCD) - the fast-sampling defaults. Leave them unless you know why you're changing them.
• Optional vae, inpaint_image + inp_strength, masked_latent, and clip_vision enable inpaint-style text placement into an existing image.

Output is a single latent (LATENT) - decode it to see the result.

How to install it

ComfyUI Manager, search ComfyUI_Anytext, or:

cd ComfyUI/custom_nodes
git clone https://github.com/zmwv823/ComfyUI_Anytext

then restart. Place the Glyph-SDXL checkpoints in custom_nodes/ComfyUI_Anytext/Image_Generation/Glyph_SDXL/checkpoints and google/byt5-small in ComfyUI/models/text_encoders (or Auto_Download). This is a heavy SDXL-scale pipeline, so budget the VRAM.

Common issues & troubleshooting

Resolution and aspect are locked. Glyph-SDXL here runs at 1024×1024, 1:1 only - the input mask and image must be square. If you're feeding a portrait or landscape mask, that's your error.

Backgrounds look worse than plain SDXL. Known limitation the author calls out. The model trades general image quality for text accuracy, and the fused pipeline doesn't match a clean SDXL checkpoint on scenery. If the background matters more than the text, this is the wrong tool.

You can't find working weights. Microsoft pulled the original Glyph-ByT5 checkpoints (private training data, pending review), and the community has spent threads just trying to relocate copies. Sourcing a functioning checkpoint is a real part of the setup cost.

Honest recommendation. For accurate multilingual text in 2026, modern base models (Flux, Qwen-Image, Ideogram-class) spell well natively and without any of this scaffolding. Glyph-ByT5 is worth a look as a curiosity or for its paragraph-layout trick, but the author's own "use the official demo instead" is good advice here.