Latent Noise Injection Sampler (CRT)

Plastic skin. Over-smoothed detail. That flat, slightly airbrushed look you get out of Flux and a few other flow-matching models at default settings. There's a known community fix for it - re-inject a bit of fresh noise partway through sampling rather than letting the whole trajectory denoise cleanly - and this node is that fix, built into the sampler itself instead of something you bolt on separately.

How it works

Under the hood this is a full custom KSampler: model, positive/negative conditioning, sampler and scheduler choice, CFG, steps, denoise, seed - the standard set. On top of that it layers two systems. First, a two-stage sigma-shaping pass (stage1_sigma_factor/stage1_sigma_start/stage1_sigma_end and the matching stage2_* fields, plus details_amount_stage1/stage2), which reshapes how aggressively detail gets resolved across two separate noise ranges - conceptually similar to the "detail daemon" style sigma tricks the community has been hand-building with third-party sampler packs. Second, and the node's namesake: at a chosen point in the sampling trajectory (injection_point), it injects a controlled amount of extra noise (injection_strength) using its own seed offset (injection_seed_offset), optionally normalized (normalize_injected_noise), and continues denoising from there. This is exactly the "use noise injection in ComfyUI" fix that's floated around as a mitigation for Flux's homogenized-face problem - this node makes it a built-in toggle instead of a separate manual step.

There's also an edit_model_flux2klein flag that enables ReferenceLatent-style conditioning specifically for edit models like Flux 2 Klein, which has largely displaced Kontext as the community's preferred local editor.

The inputs that actually matter for a first run

This node has a lot of knobs, and most of them are not what you should touch first:

• model, positive, and optional negative, latent_image - the usual sampler wiring.
• sampler_name / scheduler / cfg / steps / denoise / seed - standard KSampler fields, same behavior as anywhere else.
• enable_noise_injection (default "enable") and injection_strength (default 0.1) - the one dial worth adjusting if output looks too smooth: nudge injection_strength up a bit, or turn injection off entirely (disable) if you want a plain sampler with none of this behavior.
• Everything under stage1_*/stage2_*/details_amount_* and the rest of the injection fields - leave these at default until you understand exactly what they're doing. They're genuinely advanced sigma-manipulation controls, not settings a first pass needs.

Optional image and vae let it run as an image-to-latent sampler too. Outputs are latent and image.

How to install it

Part of CRT-Nodes - via ComfyUI Manager, search CRT-Nodes and install; manually, git clone https://github.com/PGCRT/CRT-Nodes.git into custom_nodes, pip install -r requirements.txt, restart.

Common issues & troubleshooting

Output looks noisier or grittier than a plain KSampler, not more detailed. injection_strength is too high relative to your model and step count - back it off in small increments rather than large jumps; this parameter is sensitive.

No visible effect at all. Check enable_noise_injection is actually set to "enable" - it's an enum, not a boolean, and it's easy to assume it's on by default when toggling other fields.

Trying to use it as a plain edit-model sampler and getting odd conditioning behavior. edit_model_flux2klein changes how conditioning is applied specifically for reference-based editing - leave it off unless you're actually working with an edit model like Klein that expects that ReferenceLatent-style setup.