JoyType Sampler

This is a self-contained SD 1.5 sampler that runs through HuggingFace's diffusers library rather than ComfyUI's native sampling. Its display name in the menu is "JoyType Sampler," and that's its main job - pairing an SD 1.5 diffusers checkpoint with the JoyType canny control image to render text into a picture. But the control input is optional, so it also works as a plain text-to-image sampler.

The reason it exists as a separate node is that JoyType (and the diffusers path generally) doesn't slot into Comfy's normal MODEL/CONDITIONING wiring - it wants a diffusers pipeline. This node wraps that whole pipeline into one box.

How it works

You feed it a diffusers_model (from UL_DiffusersCheckpointLoader), your prompts, and optionally a diffusers_control - which is the canny_image from the JoyType Font_Img node. The node runs the diffusers denoise loop and returns a finished IMAGE directly (not a latent - it does the VAE decode internally). Because it's diffusers, it exposes its own scheduler list (28 options, Euler a by default) and a device selector, and it manages model residency with its own keep_model_loaded flags rather than Comfy's.

The inputs and outputs that matter

• diffusers_model (Diffusers_Model) - the checkpoint, from the diffusers loader. Required.
• prompt - your main scene text. The a_prompt (appended quality terms) and n_prompt (negative) come pre-filled with sensible text-rendering defaults ("supper legible text, clear text edges…") - leave them unless you have a reason.
• steps (20), cfg (7.5), scheduler (Euler a) - standard SD 1.5 sampling values, and standard here.
• width / height (default 512) and batch_size - SD 1.5 wants ~512; push resolution too far and you get the usual SD 1.5 duplication.
• diffusers_control (optional, Diffusers_Control) - the JoyType canny hint. Connect it for text rendering; leave it off for plain text-to-image.

Output is image (IMAGE) - ready to save or post-process.

How to install it

ComfyUI Manager, search ComfyUI_Anytext, or:

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

then restart. For the JoyType use case, download the JoyType-v1-1M canny ControlNet into ComfyUI/models/controlnet and an SD 1.5 checkpoint for the diffusers loader. The diffusers library itself comes in with the pack's requirements.

Common issues & troubleshooting

It ignores my ControlNet. diffusers_control is optional, so if it's not connected the node just does text-to-image. For JoyType text rendering, wire the canny_image output of the Font_Img node into it.

Slow first run. Diffusers loads and compiles its pipeline separately from Comfy's model manager. The first generation after loading is slower; keep_model_loaded keeps it warm for subsequent runs. Turn that off if you're short on VRAM and swapping models.

Duplicated subjects / mangled composition at high res. Classic SD 1.5 behavior above ~512–768px. Keep the base resolution modest and upscale afterward rather than sampling large.

Why diffusers instead of a native KSampler? Only because JoyType ships as a diffusers pipeline. If you'd rather stay in native ComfyUI, the author notes JoyType's canny model also works with a normal Comfy KSampler + ControlNet - this node is just the batteries-included option.