ComfyUI error: Wrong latent / EmptyLatentImage for SD3

Use EmptySD3LatentImage (ComfyUI core) as in the SD3.5 example workflow.

ComfyUI error: CUDA out of memory

Stay on the FP8 all-in-one checkpoint; if using split weights, use t5xxl_fp8_e4m3fn_scaled with under 32GB system RAM (per source); try --lowvram.

SD3.5 Large FP8 text-to-image

Text-to-image with Stability AI Stable Diffusion 3.5 Large using the Comfy-Org all-in-one FP8 checkpoint (text encoders included).

Estimated VRAM

~14.935 GB — ⚠️ ESTIMATE (unverified real-world usage). Estimate, not a measured figure: weight-size floor = sd3.5_large_fp8_scaled.safetensors (14.935 GB, sizes verified on Hugging Face). Activations add more; ComfyUI offloading can reduce it. Source, checked 2026-10-14.

Models required

sd3.5_large_fp8_scaled.safetensors — 14.935 GB ✅ verified 2026-10-14 (listing) · sha256 5ad94d6f951556b1ab6b75930fd4effbafaf3130fe9df440e7f2d05a220dd1be → models/checkpoints

Tested GPUs

No verified test reports yet.

Nodes

CheckpointLoaderSimple CLIPTextEncode EmptySD3LatentImage KSampler VAEDecode SaveImage

Common errors & fixes

Wrong latent / EmptyLatentImage for SD3: Use EmptySD3LatentImage (ComfyUI core) as in the SD3.5 example workflow. ✅ per source
CUDA out of memory: Stay on the FP8 all-in-one checkpoint; if using split weights, use t5xxl_fp8_e4m3fn_scaled with under 32GB system RAM (per source); try --lowvram. ✅ per source

Low-VRAM tips

⚠️ General guidance, not individually verified.

Use the FP8 all-in-one checkpoint (already used here)
If using split weights, prefer t5xxl_fp8_e4m3fn_scaled when under 32GB system RAM (per ComfyUI_examples)
Launch with --lowvram if needed

Source & credit

Workflow by comfyanonymous (ComfyUI_examples) / Stability AI. Get it from the original: https://comfyanonymous.github.io/ComfyUI_examples/sd3/. License: Stability AI Community License.