🖼️PixArtAlpha Generation

The ModelLoader downloads the model, the Styler writes the prompt, and this is the node that actually makes the picture. It's the pack's sampler: feed it a pipe, a positive and negative prompt, and it runs the diffusion loop for N steps and hands back a standard IMAGE you can preview or save. If you've used any text-to-image node in ComfyUI, this will feel familiar - it's the KSampler-shaped piece of a workflow that otherwise looks nothing like a normal checkpoint setup.

How it works

Under the hood it picks a scheduler, seeds a torch generator on the same device as the model, runs the diffusers pipeline, and converts the PIL output to a ComfyUI IMAGE tensor. The scheduler choice is the interesting bit: DPM-Solver uses diffusers' DPMSolverMultistepScheduler, while SA-Solver uses a scheduler ported into the pack itself. Swapping between them is cheap - it only replaces the scheduler object, never reloads the model - so you can experiment freely.

The inputs that matter

• steps - default 20. The official PixArt recipe is 20–25, and 20 genuinely is the sweet spot. Pushing past ~30 mostly wastes time.
• guidance_scale - default 4.5. The author's rule: keep 4.5 on DPM-Solver, drop to ~3.0 on SA-Solver. This is the knob people over-turn; PixArt goes washed-out and crunchy at high CFG.
• width / height - default 1024×1024, multiples of 32. PixArt Alpha is resolution-picky (it trained at 1024 and 512), so 1024×1024 is the reliable recipe. Non-native sizes are where people get broken-looking images.
• schedule - the DPM-Solver / SA-Solver toggle. SA-Solver reaches similar quality in fewer steps, which is why it wants a lower guidance value.
• seed - 0 means a fresh seed each run; set a value to reproduce an image exactly.

The positive and negative inputs are STRING fields with force-input turned on, meaning there's no text box by default - they're meant to be wired in. In the pack's own workflow those wires come from the Styler node, but any text source (or a converted text box) works.

Output and troubleshooting

The single output is IMAGE, the standard ComfyUI tensor - wire it to PreviewImage or SaveImage. A few things to watch:

• Washed-out or overbaked results - your guidance is too high for the scheduler you picked. Follow the 4.5 / 3.0 split above.
• Garbled images - you're off the native 1024×1024 resolution, or you fought the resolution step and ended up at something the model never trained on.
• Slow first generation - the first run after loading is dominated by the T5 text encoder; that's normal, not a hang.

This node deliberately exposes no CFG rescale, no refiner, no advanced sampler settings - it's a minimal wrapper, so don't go hunting for the full KSampler arsenal here. Keep it simple, and it's one of the more reliable "small model, surprisingly good prompt adherence" experiences left in ComfyUI after Flux took over the headlines.