RunningHub OneReward Loader

This is the node that explains why your disk suddenly lost 30 GB. One dropdown, one output, zero sampling - the RunningHub OneReward Loader exists to fetch ByteDance's OneReward fine-tune, drop it on top of the FLUX.1 Fill base model, and hand you a ready-to-use pipeline. You wire that pipeline into the pack's Sampler or Eraser and never touch it again.

The model is the whole story. OneReward is an RLHF fine-tune of FLUX.1-Fill-dev that ByteDance trained across image fill, extend, and object removal at once, using Qwen2.5-VL as a generative reward model. In the 2026 inpainting argument - FLUX.1 Fill dev vs Flux 2 Klein vs this - OneReward is one of the two still-standing dedicated mask models, and its defenders claim it clears the closed FLUX Fill Pro on its own tasks. The weights are CC BY-NC 4.0, the code Apache 2.0, so it's personal-use only. Not a downside for most of you; worth knowing before you build a product on it.

Mechanically it's simple and worth understanding once. The loader grabs a transformer from models/OneReward/<model>, pulls the full FLUX.1-Fill-dev directory from models/black-forest-labs/FLUX.1-Fill-dev, and swaps the OneReward transformer into the base, all in bfloat16. The result is wrapped in the pack's FluxFillCFGPipeline - a subclass of diffusers' FluxFillPipeline that adds the "true CFG" branch the Sampler uses. Memory handling is automatic: if your GPU has under 30 GB of VRAM it int8-quantizes the transformer with optimum.quanto and enables CPU offload; at 30 GB and up it just moves everything to CUDA (the source literally prints "run in 48g mode"). The README says consumer cards around 16 GB work - true, but the threshold in the code is 30 GB, so a 24 GB card gets quantized too, and loading takes a while.

The one input you set:

• model - the dropdown. flux.1-fill-dev-OneReward-transformer (default) or flux.1-fill-dev-OneRewardDynamic-transformer. The standard one is the safe pick; the Dynamic variant is the optional extra ByteDance released, and you only need it if you specifically want its behavior.

The output is OneReward Pipeline, a custom RHOneRewardPipeline type that only the Sampler and Eraser accept. Don't expect to plug it into anything else - it's a closed type, and that's fine because those two nodes are the whole point.

Install is the README's way, with one trap:

cd ComfyUI/custom_nodes
git clone https://github.com/HM-RunningHub/ComfyUI_RH_OneReward

or just search "ComfyUI OneReward" in ComfyUI Manager. Then the model downloads - roughly 24 GB for FLUX.1-Fill-dev into models/black-forest-labs/FLUX.1-Fill-dev/, and ~6 GB for the OneReward transformer into models/OneReward/flux.1-fill-dev-OneReward-transformer/:

huggingface-cli download black-forest-labs/FLUX.1-Fill-dev --local-dir models/black-forest-labs/FLUX.1-Fill-dev
huggingface-cli download black-forest-labs/flux.1-fill-dev-OneReward-transformer --local-dir models/OneReward/flux.1-fill-dev-OneReward-transformer

Two gotchas the README hides. First, it tells you to pip install -r requirements.txt, but there is no requirements.txt in the repo - that command fails. What the code actually imports is diffusers (new enough to have FluxFillPipeline, so mid-2024 or later) and, only in the low-VRAM path, optimum-quanto. Install those yourself. Second, the model folders must be named exactly as the loader builds them - models/OneReward/flux.1-fill-dev-OneReward-transformer, not models/OneReward/whatever-else. If the loader errors at "no such file or directory," it's almost always that.

Restart ComfyUI after installing, let it do the one-time quantization dance on first load, and you're ready to point the Sampler at your mask.