HiDream I1 Fast
Generate images quickly with HiDream I1 Fast - Lightweight version with 16 inference steps, ideal for rapid previews on lower-end hardware.
About
HiDream I1 Fast is a lightweight text-to-image workflow tuned for rapid previews and low-VRAM GPUs. It pairs the HiDream I1 Fast FP8 UNet (loaded with UNETLoader) with SD3-style sampling (ModelSamplingSD3) and runs just 16 inference steps in KSampler. Prompts are handled through QuadrupleCLIPLoader and CLIPTextEncode, while EmptySD3LatentImage creates the starting latent at your chosen resolution. The result is a fast, responsive pipeline that’s ideal for exploring ideas and iterating on compositions without long waits.
What you provide
- Reading the workflow
Runs in ComfyUI. This workflow can't run on this page. Open it in Comfy Cloud with everything loaded, or download it for your own ComfyUI.
Try in CloudQuestions & answers
Keep steps at 16, lower resolution in EmptySD3LatentImage, reduce batch size to 1, and ensure you’re using the FP8 model (hidream_i1_fast_fp8.safetensors). Larger images and higher steps raise memory use.
The ModelSamplingSD3 shift balances conditioning across the multiple text encoders used by HiDream/SD3-style models. Start near 3.0 (official guidance) and adjust slightly if prompts feel over/under-strong.
The workflow is sampler-agnostic. Per HiDream’s notes, uni_pc with a simple scheduler is a reliable default. If you need different behavior, try Euler or DPM++ variants and compare results at the same step count.
Increase steps moderately (e.g., 20–24), refine your positive/negative prompts, tweak shift (±0.5), and consider a slightly higher resolution. Also set a fixed seed in KSampler for reproducible tuning.