DDRK Omega Lite

Where this sits

There's a version of sampler confusion that isn't "which sampler do I pick" but "I already know my numbers". You've got the model card, it says 25 steps at CFG 6, or 8 steps at CFG 1, and you're not interested in relitigating that. What you don't have is an informed opinion on integrator order, step placement or whether a little sharpening at the end is a good idea.

That's the gap DDRK Omega Lite fills. It takes your steps and your CFG as given and maps everything else - integrator, error-based step adaptation, and optionally the character enhancers - onto combinations the pack's author validated against the full control node. The mapping is verified with torch.equal against the equivalent full-node settings, so the presets aren't approximations of the big node, they are the big node with the dials preset. It's a thin wrapper over the same sampler code, not a second implementation.

The two dials

quality is where all of it happens, and it means different things by family.

On flow-matching models (Flux, SD3, Qwen, Chroma, Anima, Wan) all three levels run the same second-order HC2 integrator at one model call per step. fast is plain HC2, balanced adds step adaptation at 0.10, and best adds a second pass - the finished latent is upscaled to 1.33x and partly re-sampled at denoise 0.35 for about 40% of your step count. So on flow matching, quality is not a speed control. The speed control is the steps widget, exactly as it should be; the dial trades numerical aggressiveness until you reach best, which is the one level that costs real time - about 1.7x in the author's measurements, plus VRAM for the bigger canvas.

On EDM models (SD 1.5, SDXL and their finetunes) the levels are integrator cost, plain: fast is Euler, balanced is Heun at two model calls per step, best is RK4 at four. Compare those at equal model calls, not equal steps, or fast looks mysteriously fine.

character is the style layer, and neutral - the default - turns all of it off. sharp applies a 0.12 final-step sharpen, which is a flow-matching-only feature; on SDXL it's bit-identical to neutral because sharpening is disabled for that family by design. smooth turns on SABER spatial fusion at 0.10, which is a deliberate smoothing control that measurably narrows dynamic range. It's exposed rather than defaulted for that reason: if you pick it, you picked it.

Inputs and output

model, positive, negative, latent_image, seed, steps (default 20), cfg (default 5), then the two enum dials. One output: LATENT.

Two absences are worth knowing before you wire it up. There's no denoise - it samples at 1.0, so Lite is text-to-image only; for img2img use DDRK Omega Auto, which does expose denoise, or the pack's advanced Unified node. And there's no settings output and no console summary, so Lite is silent about what it chose. If you want to see the decision, that's the Auto node.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/HVOSTOVSKY/DDRK-Omega-Sampler

Restart ComfyUI; the node appears under sampling → DDRK Omega. Despite the version churn in this pack there are no Python dependencies - it's pure ComfyUI plumbing - so installation is not where things go wrong. ComfyUI Manager works too: search the pack title.

Where people get burned

Below 7 steps the sampler overrides your integrator to Euler unless you're on HC2 or HC3, and it does that regardless of any other setting, printing a line to the console when it happens. On Lite that bites the EDM side: balanced and best at 5-6 steps are really running Euler, so the preset you paid for isn't the preset you got. Similarly, on EDM at 10 steps or fewer RK4 is quietly downgraded to Heun. Neither is a bug - high-order methods need enough steps to pay for their extra model calls - but the console message is the only place it's announced.

Turbo and Lightning LoRAs are the other trap. Lite has no few-step mode because Lite has no model_type input: the node can't see that a LoRA changed your model's character. Set steps to 4-8 and cfg to 1 yourself. And keep CFG honest on those models - guidance-distilled checkpoints double their render time above CFG 1 and cook long before it looks like a quality gain. Same story for the negative prompt: at CFG 1 it has nothing to steer, so write constraints as presence instead.

Two smaller ones. cfg floors at 1.0 here, so if a workflow of yours wanted 0, use the Unified node. And smooth narrowing dynamic range is documented behaviour, not a bug - if your flat-shaded or pixel-art output gets mushy, that's the dial to turn back to neutral.