ChromaRadianceOptions

ChromaRadianceOptions is the "in case you need it" node for Chroma Radiance, LodestoneRock's pixel-space image model. Chroma is the community's Apache-2.0 answer to Flux licensing - the 8.9B model everyone points at for uncensored work on a modern base. Chroma Radiance is the follow-on that went pixel-space (no VAE, direct pixel tokens), continuing Lodestone's one-man experiments in the no-latent architecture direction. This node exists to expose knobs on that model that most people will never need, which is why its description is honest: "Allows setting advanced options for the Chroma Radiance model."

There's a real chance you search for this node because a workflow loaded with a red "missing" badge and you have no idea what it does. Good news: if you're not running a Chroma Radiance checkpoint, you don't need it at all. If you are, here's what's behind the door.

How it works

The node patches the model with a wrapper that injects per-step options into the model's forward pass, but only within a sigma window. The two sigma inputs - start_sigma (default 1.0) and end_sigma (default 0.0) - define when the options are in effect, in the model's sigma schedule. During that window it stuffs the settings into the transformer options so the Radiance model's internal code picks them up; outside it, it's a pass-through. If you never change any option, the node returns your model untouched.

The two options that actually do something:

• nerf_tile_size (default -1) overrides the model's internal tile size. -1 means "use the default (32)"; 0 disables tiling entirely - which the tooltip warns may require a lot of VRAM. Chroma Radiance samples in NeRF-style tiles to keep memory sane at high resolution, so this is a memory/quality trade.
• force_sequential_txt_ids (default off) is a compatibility shim for a specific range of checkpoints (trained between 2026-05-22 and 2026-06-01) that use sequential text token IDs but forgot to mark it in their state dict. You will almost certainly never need it.

Inputs and what to set

• model - the Chroma Radiance model.
• preserve_wrapper - default on, "generally should be left enabled." It delegates to an existing model wrapper if one's already there so the node doesn't clobber another patch in the chain.
• start_sigma / end_sigma - when the options apply. Defaults cover the whole schedule.
• nerf_tile_size - the one knob you might actually use, and only if you're chasing quality at the cost of VRAM.
• force_sequential_txt_ids - the checkpoint-shim flag. Leave it off.

Output: the patched MODEL. Ships with ComfyUI core.

Common issues & troubleshooting

Missing-node errors on this are usually harmless. If you loaded a Radiance workflow without this node, the only thing you lose is whatever advanced option the author set - usually nothing. Wire in a passthrough or just delete it if the workflow errors.

Turning off tiling OOMs you. nerf_tile_size 0 is the "full image in one pass" mode and the tooltip's warning is not a suggestion. Only try it on a card with real headroom, and expect the memory graph to spike.

It's model-specific by construction. The options only matter if the model reads chroma_radiance_options from its transformer options. On any other checkpoint this node is a no-op passthrough.