🤖 Leon Stable Diffusion 3.5 API

Stable Diffusion 3.5 is a strange chapter in the lineage: the MMDiT architecture was genuinely novel, but the launch cycle was so rocky - broken base releases, a CivitAI ban, and a finetune ecosystem that never formed - that it's remembered as the family that stumbled. Running SD 3.5 locally is entirely possible, but it's a heavy 8B-class model (2B for Medium) that needs the right stack to be pleasant. Leon_StableDiffusion_35_API_Node sidesteps the whole local question: it calls sd3.5-large, sd3.5-large-turbo, or sd3.5-medium through the HyprLab aggregator, so you can point ComfyUI at Stability's MMDiT without the weights, the VRAM, or the setup.

Reach for it when you want SD 3.5's strengths - strong prompt adherence, good text rendering in the 3.5 generation - but your machine (or your patience for 8B-model management) says no. It's also a decent way to A/B the three sizes before deciding whether any of them deserve a local install.

How it works

The pack's Stable Diffusion nodes share their own StableDiffusionImageGenerationNodeBase, which is functionally the same as the HyprLab base used elsewhere: build a payload, POST to https://api.hyprlab.io/v1/images/generations with a Bearer token, retry with exponential backoff (five attempts), decode to an RGBA tensor. The SD-specific payload fields here are negative_prompt and aspect_ratio, and the prompt field explicitly supports (word:weight) weighting syntax per the tooltip - a taste of the classic SD control you give up on most API nodes.

Inputs that matter

• model - sd3.5-large (default), sd3.5-large-turbo, or sd3.5-medium. Turbo trades some fidelity for speed; Medium is the light 2B. If you're cost-sensitive and the prompt isn't demanding, Medium gets you 80% there for a fraction of the compute.
• negative_prompt - the classic SD lever most API wrappers forget. Fill it with what you don't want.
• aspect_ratio - nine options (16:9, 1:1, 21:9, 2:3, 3:2, 4:5, 5:4, 9:16, 9:21).

Outputs: image, image_url, seed. The seed field here caps at 4,294,967,294, matching the SD3 API's range.

Installing

Same pack, same two ways - ComfyUI Manager (search Leon) or clone:

cd ComfyUI/custom_nodes/
git clone https://github.com/l3ony2k/comfyui-leon-nodes comfyui-leon-nodes
pip install -r requirements.txt

Restart, paste a HyprLab key. No model downloads, no VRAM.

Common issues

The pack-wide classics: left-at-default api_key, and reading the console for the exact 4xx body. SD3-specific, remember that this family historically cares about prompt formatting - the API's (word:weight) syntax is real, and if your prompt has stray parens, the parser can behave oddly, so keep the syntax deliberate. And note the seed ceiling: pasting in a seed above ~4.29 billion will fail validation, which is a different error from the ones you'll see on the pack's other nodes.