Chroma Unified Model Loader

Quick heads-up before anything else: the pack's own README opens with "DEPRECATED! Chroma Manual" and tells you to use ComfyUI's default Chroma workflow instead of this repo. That covers this node. So if you're staring at a workflow that uses ChromaDiffusionLoader and wondering why it's not in your Manager search results by default, or why nobody talks about it - that's why. It still works if you install the pack, but it's not the recommended path for new Chroma setups anymore.

A little history explains why it exists at all. ComfyUI_FluxMod comes from LodestoneRock, the same solo developer who later trained Chroma itself - an 8.9B, Apache 2.0, uncensored rework of Flux Schnell. This node predates that: it's a one-stop loader that reads a Chroma checkpoint and hands you back a ready-to-sample MODEL with quantization baked in, built while Chroma was still an early, moving target and the ecosystem didn't have first-class support for it yet. Once ComfyUI added Chroma natively - comfyanonymous's team implemented it directly - that first-class support did everything this node did, using the plain UNETLoader, so the custom loader became redundant.

What it does

Two inputs. unet_name picks your model file - the tooltip confirms it reads both .safetensors and .GGUF files, the latter only if ComfyUI-GGUF is also installed. quant_mode picks how the weights get cast at load time: bf16 (full quality, wants 24GB+), float8_e4m3fn (8-bit, the recommended compromise for 12-16GB cards), or float8_e5m2 (also 8-bit, but the pack's own docs mark it not recommended). One output: model, which you wire straight into your sampler.

If that sounds exactly like what a normal ComfyUI UNETLoader plus a weight_dtype dropdown does - that's basically the point. Native ComfyUI now covers the same ground for Chroma without a custom node in the mix.

Installing it (if you need it)

Via ComfyUI Manager: search "ComfyUI_FluxMod" in the Custom Nodes Manager, install, restart. Manually:

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

Then restart ComfyUI and refresh your browser. No extra Python dependencies beyond ComfyUI-GGUF if you want GGUF checkpoints to show up in unet_name.

What you actually need for Chroma today, per the README's own recommendation: a Chroma checkpoint in ComfyUI/models/diffusion_models, a T5-XXL text encoder (fp16 or fp8, either works) in models/clip, and the FLUX VAE in models/vae. Load the official example workflow from comfyanonymous's ComfyUI_examples/chroma page, and you're generating with the stock UNETLoader - no FluxMod install required.

Common issues & troubleshooting

You can't find this node when searching Manager for "Chroma." It ships inside ComfyUI_FluxMod, not as its own listing - search for the pack name, not the node name.

Your Chroma outputs look worse than expected. Before touching this node, check the basics the KB flags repeatedly for Chroma: it wants ridiculously verbose, descriptive prompts - short prompts produce unreliable results - and it's genuinely slow unquantized on anything under 16GB. float8_e4m3fn is the sane middle ground if you're VRAM-constrained; skip float8_e5m2, the pack authors themselves don't recommend it.

You're mixing this loader with FluxMod's other nodes and things break. Don't. The README is explicit that everything below its deprecated Chroma section is a different feature - modulation-layer quantization for stock Flux Dev/Schnell, not Chroma. KSamplerMod and FluxModSamplerWrapper belong to that other half of the pack, not to a Chroma workflow built around this loader.

Bottom line: if you're starting fresh, skip this node and use the native ComfyUI Chroma workflow. This one's worth knowing about mainly so an old workflow JSON doesn't confuse you.