Flux Chaos Guidance

Flux Dev's guidance scalar is where its obedience lives. Because the model is guidance-distilled, that single number carries all the prompt-adherence that SDXL spread across CFG - and the community consensus has long been that the 3.5 default is too high, with the sweet spot for loose, natural output sitting around 1.8–2. FluxChaosGuidance is the pack's answer: a thin node that sets a low guidance value on your conditioning, with an optional per-run jitter so each queue gets slightly different looseness without you touching anything.

It's the least exotic node in the ComfyUI-FluxHallucinate pack, and honestly the most habit-forming.

How it works

This is a CONDITIONING in, CONDITIONING out transform that writes one thing: the guidance field in the conditioning metadata - the exact same field ComfyUI's built-in FluxGuidance node writes. Nothing about the embeddings or the model is touched. That's the whole mechanism, and it's why the effect is so predictable: low guidance relaxes the distilled adherence, and output gets dreamier, looser, and more willing to surprise you.

The interesting part is jitter. When it's above zero, the node picks a random offset in the range ±jitter each queue and adds it to your base guidance (clamped at 0). So with guidance 1.5 and jitter 0.5, any given run lands between 1.0 and 2.0 - a different dose of looseness every time, seeded so you can pin it if you want. It's per-run, not per-step, which keeps the sampler sane.

The inputs

• guidance (default 1.8): the base value. The README suggests trying 1.0–2.5. Below ~1.5 you start dropping literal prompt concepts; that's the point, just know it's coming.
• jitter (default 0): per-run random variation around the base. 0.3–0.5 is a good starting wobble.
• seed: controls the jitter draw, so a fixed seed gives you the same offset every time.

The output is a single CONDITIONING, wired exactly like the input.

Wiring it up

On the conditioning line, anywhere between CLIP Text Encode and the sampler/guider - typically after FluxHallucinate in the pack's recommended recipe. It just needs to be upstream of the sampler so the guidance value rides along in the conditioning.

One hard rule from the README: don't stack it with a separate FluxGuidance node. Both write the same field, and last one wins, so stacking silently discards one of them. Pick this node (for the jitter) or the built-in one (for a fixed value), not both.

Installing

Part of the same single-file pack:

cd ComfyUI/custom_nodes
git clone https://github.com/isam2024/ComfyUI-FluxHallucinate

Restart ComfyUI afterward. No dependencies beyond torch, no model downloads. If ComfyUI Manager lists "ComfyUI-FluxHallucinate" you can install it there instead.

Common issues

• "My prompt stopped mattering" - you went too low. This node trades adherence for looseness by design; if whole concepts vanish, bump guidance back up or reduce jitter.
• "Different image every run even with the same seed" - that's jitter doing its job. Set jitter to 0 for a fixed value, or accept it: this is the node that's supposed to vary.
• "It does nothing next to my other guidance node" - see above, last one wins. Remove the duplicate.
• You might not need it at all - if you just want a flat low guidance, the built-in FluxGuidance node does it with fewer clicks. This node earns its keep specifically when you want the jitter, or when you're building the full chaos recipe from the pack README: low guidance + degraded conditioning + per-step noise + an ancestral sampler, one layer at a time.