ComfyUI lesson 8: Run modern models on modest VRAM
Choose quantization, offload, tiling, or an accelerated custom node deliberately.
Listen to this article
On-device voiceUses the voice built into your browser; no article text leaves this page.
Audio narration is not supported by this browser.
Low-VRAM work is a set of tradeoffs, not one magic flag. First determine whether weights, activations, attention, or VAE decoding causes the peak; then choose the least damaging intervention.
FP8, GGUF, and offload in plain language
| Technique | Saves | Tradeoff |
|---|---|---|
| FP8 weights | Roughly half vs FP16/BF16 weights | Hardware/node support and some precision loss |
| GGUF quantization | Multiple bit-width choices, often CPU-friendly | Extra dequantization; node-specific support |
| CPU offload | GPU memory | PCIe transfer and latency |
| Attention slicing | Activation memory | More kernel work |
| VAE tiling | Decode/encode memory | Possible tile seams, slower output |
| Smaller canvas/batch | Activations everywhere | Less resolution or throughput |
Quantized files are model variants, not interchangeable extensions. Use the loader made for that format. If a workflow expects an FP8 diffusion model plus separate text encoders and VAE, a monolithic checkpoint loader is the wrong node.
The accelerated lesson workflow
The gallery's CuteDSL Z-Image Turbo graph uses the open CuteZImageTurbo custom node. It prefers the installed CuteDSL path and falls back to the stock Diffusers Z-Image pipeline. Eight distilled steps make it a practical iteration workflow; model provenance and the custom-node source are linked on the card.
CuteDSL “latent teleportation”—reusing or transporting intermediate latent state between nearby prompt states—is an active experimental direction in this codebase, not a stable public workflow contract today. The tutorial intentionally does not fake a runnable teleport flag. When it lands, it must ship with determinism bounds, fallback behavior, and before/after visualbench evidence. Until then, use the tested accelerated Z-Image workflow and normal seed-controlled batches.
Start memory debugging at batch one and 768 or 1024 pixels. Change one optimization at a time and record peak VRAM plus generation time. A workflow that fits but takes ten times longer may not be the win you wanted.
Run the lesson, do not just read it
- Open the tested workflow in Comfy Studio
- Download the exact API-format JSON
- CLI:
app comfy download cute-zimage-turbo
The gallery records model sources, minimum VRAM, custom-node requirements, and verification state beside the graph. That is the portable unit for this series.
Series progress: lesson 8 of 10. Previous: Control composition with ControlNet Next: Automate batches, prompt lists, REST, CLI, and MCP