Hunyuan 3 Generate (Low VRAM Budget)

"Hunyuan 3 Generate (Low VRAM Budget)" is the generator end of the pack's recommended 24–32GB pairing: it's the Low VRAM generator plus a gpu_budget_gb override and memory telemetry, designed to run against the NF4 Low VRAM+ loader. The README's verified reference workflow - loader → (optional rewriter) → this node → save - produced a 1600×1600 render on a 24GB rig without tripping the bitsandbytes validation that used to plague NF4 on small cards. If you have a 24GB GPU and want HunyuanImage-3.0 at all, this is the node you'll live in.

Like the plain Low VRAM generator, it's built for quantized NF4/INT8 models loaded with a device map: it skips the conflicting CPU-offload calls and works with accelerate's offloading instead of against it. The Budget variant adds the two things production workflows want - a way to override the memory assumption for a single run, and telemetry in the status output.

The inputs that matter

The usual generator controls are all here (model, prompt, seed, steps default 40, resolution, guidance_scale, offload_mode, post_action), plus the optional prompt-rewrite block and two things that make this node what it is:

• gpu_budget_gb - override the loader's recorded GPU budget for this run (default -1 = use the loader's value). The README gives the concrete use: set it to assume a different baseline (say 65GB) so smart-mode offload math changes, without reloading or remapping the model. Leave it at -1 unless you know why you're changing it.
• post_action defaults to soft_unload_to_cpu - the right default on a card where you'll want VRAM back for downstream nodes.

offload_mode has the same honest tooltip as the Low VRAM node: for NF4/INT8 it's ignored, because the loader's device map already owns offloading. Set always and you'll double-offload and stall. Outputs are image, rewritten_prompt, status (now with memory telemetry), and trigger.

The workflow it belongs to

Hunyuan 3 Loader (NF4 Low VRAM+) → Hunyuan 3 Generate (Low VRAM Budget) → Save/Display

The loader's gpu_memory_target_gb (18–20GB on a 24GB card) is what this node's smart defaults assume - keep them in sync and it stays out of trouble. Swap in a plain Low VRAM generator if you don't need the budget override or telemetry; they're compatible, but this variant is the fuller-featured choice.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Comfy_HunyuanImage3
cd Comfy_HunyuanImage3
pip install -r requirements.txt

Restart ComfyUI (or ComfyUI Manager, search "Comfy_HunyuanImage3"), and have the NF4 weights in ComfyUI/models/:

huggingface-cli download EricRollei/HunyuanImage-3-NF4-v2 --local-dir HunyuanImage-3-NF4

Reality check

All the Low VRAM caveats apply: this node removes crashes, not physics. On 24GB cards the offloaded weights stream over PCIe, so generations are slow - community testers who ran NF4 on a 4090 called it functional but "not worth the time" for serious iteration, and a 96GB card runs circles around it. The budget slider gives you fine-grained control, but it can't manufacture bandwidth. If you've got a 48GB+ card, use the base Generate node and keep this one for a future upgrade scenario.