FluxMod Unified Model Loader

This is the main event of ComfyUI_FluxMod - and to be clear up front, it's a different feature from the pack's Chroma nodes. The README splits sharply: everything under its "Chroma Manual" heading is deprecated, but this loader and its siblings (KSamplerMod, FluxModSamplerWrapper, SkipLayerForward) are the pack's actual, still-current purpose - a modulation-layer addon that shrinks Flux Dev or Flux Schnell from 12B parameters down to roughly 8.8B, letting them run on more consumer-friendly hardware without a big quality hit.

It's the same core trick LodestoneRock later used at full retraining scale to build Chroma - swap out Flux's oversized modulation layers for a smaller replacement. The difference is that this pack applies it as an addon on top of an existing, official Flux Dev/Schnell checkpoint, while Chroma is a from-scratch retrain that bakes the change in permanently. Think of FluxMod as the earlier, lighter-weight version of that idea, and Chroma as what happened when it got scaled all the way up.

One honest note: this didn't land as a big community moment when it shipped - the launch thread got a handful of comments, several of them people asking what it offers over quantizing Flux the normal way (fp8, GGUF Q8). That's a fair question and there isn't a single crisp answer; the modulation swap and quantization are solving overlapping but not identical problems, and if you're already happy running fp8 or GGUF Flux, you may not need this at all.

What it does

Three inputs. unet_name is your base Flux checkpoint - the tooltip confirms it reads .safetensors and, if ComfyUI-GGUF is installed, .GGUF files too. guidance_name is the modulation addon itself - download universal_modulator.safetensors from lodestone-horizon/flux-essence on HuggingFace and it'll show up here. quant_mode picks how the combined model gets cast: bf16 (best quality, wants 24GB+ VRAM), float8_e4m3fn (8-bit, the recommended balance for 12-16GB cards), or float8_e5m2 (also 8-bit, but the pack's own quantization guide marks it not recommended). Output is a single model you wire into your sampler.

The one rule that trips people up: if you pick either float8 mode here, you cannot use a stock ComfyUI KSampler downstream - you need KSamplerMod, or FluxModSamplerWrapper if you're building an advanced SamplerCustom graph. That requirement disappears in bf16 mode, or if you launch ComfyUI with --fast.

How to install it

Via ComfyUI Manager: search "ComfyUI_FluxMod" in Custom Nodes Manager, install, restart. Via the Comfy Registry CLI: comfy node registry-install comfyui_fluxmod. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/lodestone-rock/ComfyUI_FluxMod.git

Restart ComfyUI and refresh your browser. You need the base Flux Dev or Flux Schnell checkpoint (official BFL weights - the README says third-party finetunes "may not work") plus universal_modulator.safetensors from lodestone-horizon/flux-essence, and ComfyUI-GGUF installed separately if you want GGUF checkpoints to show up in unet_name.

Common issues & troubleshooting

The README's own node table calls this "FluxModCheckpointLoader," but you can't find that name in your node search. The class shipped under this name - search for "FluxMod Unified Model Loader" or FluxModDiffusionLoader, not the older name from the docs.

Sampling produces garbage or errors after loading in float8 mode. Almost always the sampler mismatch above - swap the stock KSampler for KSamplerMod.

Outputs look basically identical to just running Flux normally, minus the VRAM savings. That tracks with the pack's own pitch: it's targeting hardware fit, not a quality upgrade. If VRAM isn't your constraint, there's no strong reason to add this over plain Flux.

Unsure if Flux LoRAs behave correctly on top of the patched model. This came up in the community at launch without a firm answer either way - nobody's published solid compatibility testing. Treat it as untested territory rather than assuming it just works.