RUM FLUX.2 Diffusers Euler Sampler

The shortest input list in the pack: RUMFlux2DiffusersEulerSampler has no inputs at all. You drop it in, and it hands back a SAMPLER that reproduces how diffusers' Flux2KleinPipeline steps with Euler - raw noise prediction, then x = x + dt * noise. It's the sampling half of RUM's "match the reference pipeline" project, and it's the only sampler the pack's edit path will accept.

Don't mistake the empty inputs for a generic Euler. This one is a specific Euler: it asks the model for raw noise predictions (via the rum_diffusers_return_raw_noise model option), steps with the diffusers dt convention, and handles the denoise/cleanup in the reference's order. Stock samplers estimate the denoised latent differently, which is precisely why the pack needed its own.

Why you must use it for the edit workflow

The killer feature hides in the CFG guider. When RUMFlux2DiffusersCFGuider gets reference_latents (the edit path - "change clothes to school uniform" on a reference image), the reference tokens get appended to the conditioning during sampling. That injection only works on this sampler's raw-noise path. The changelog is explicit about the failure mode: connect the guider to a regular sampler with reference_latents attached and the node raises a clear error pointing at this node instead of dying inside some attention concat. That error is a gift - newer versions of other packs would just crash.

Usage

RUMFlux2DiffusersEulerSampler ──sampler──▶ SamplerCustomAdvanced
RUMFlux2DiffusersScheduler ────sigmas───▶ SamplerCustomAdvanced
RUMFlux2DiffusersNoise ────────noise────▶ SamplerCustomAdvanced
RUMFlux2DiffusersCFGuider ────guider────▶ SamplerCustomAdvanced

That's the whole sampling chain in both shipped workflows. The sampler also refuses denoise_mask - full-image sampling only, so regional/inpaint tricks aren't supported on this path.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Rinne414/ComfyUI-RUM
cd ComfyUI-RUM
pip install -r requirements.txt
python scripts/download_models.py --comfy-root /path/to/ComfyUI --include-teacher-clip

Restart. Requires ComfyUI v0.26.2+ - the sampler leans on SamplerCustomAdvanced and nested-latent support.

The honest take

If you're on the plain native path (not chasing reference-exact results), you don't need this node at all - a normal SamplerCustomAdvanced with a standard sampler works fine and the pack's earlier versions shipped exactly that. Reach for this one when you're running the diffusers-match workflows or doing the edit path, where it's not optional. It's a narrow, well-built tool: zero configuration, one job, and it fails loudly when you use it wrong.