DiffuseHighSampler

The standard way to get a big image is the highres-fix loop: generate at native resolution, upscale, then run a few low-denoise steps. It works, but it's two stages bolted together, and past 2x the second pass starts inventing second heads. DiffuseHighSampler is the alternative that does the whole climb inside a single sampler node, and it's mostly model-agnostic - no checkpoint patching, no per-model fiddling.

DiffuseHigh is a 2024 research method from the group behind the yhyun225/DiffuseHigh repo. Blepping's "jank" port implements it as a custom sampler rather than a model patch, which is the clever bit. Deep Shrink and HiDiffusion literally patch the model - that mostly restricts them to SD1.5/SDXL and means enabling the effect changes your seed. A sampler sits upstream of the model, so jank DiffuseHigh runs on Flux and other rectified-flow models too, and you can dial the effect on and off without breaking reproducibility.

How it actually works

Each iteration is upscale → add noise back → run some steps at low denoise (~0.3–0.5). The twist is structure guidance. For the first guidance_steps of every iteration, the sampler mixes in guidance derived from a reference image (the previous pass, or an image you feed in). The reference gets split with wavelets - that's what the pywavelets and pytorch-wavelets dependencies are for - its low-frequency parts carry the composition so the model can't wander into duplicated anatomy, while the sharpened high-frequency detail comes from the denoised output. The model stays on script because it's being told "this is the structure, stay inside it," which is also why you can run these steps at a higher denoise than a plain highres fix would survive.

The inputs that actually matter

Start here and touch almost nothing else:

• iterations + scale_factor - the resolution ladder. 512×512, scale 2, three iterations gets you to 4096×4096. The tooltip's warning is not rhetorical: this adds up fast.
• guidance_mode - image matches the original method but needs a VAE encode/decode per guidance step. latent is about twice as fast; the author recommends it, and you can drop guidance_factor a bit to compensate.
• guidance_steps, guidance_factor, fadeout_factor - how long and how hard guidance applies. Fadeout eases the factor down per step (1.00, 0.75, 0.50, 0.25 for four steps), which reads smoother.
• renoise_factor - 1.0 is the normal amount of noise. Bump to ~1.02 for slightly more detail.
• vae_mode - normal, tiled, taesd (fast/cheap but needs the TAESD encoders in models/vae_approx), or tiled_diffusion if you have ComfyUI-TiledDiffusion.

The optional inputs are where the power hides: highres_sigmas (keep it around 0.3–0.5 denoise - plugging raw Karras sigmas in is a classic way to get garbage), sampler, reference_image_opt (feed your own seed image and skip the low-res pre-pass), upscale_model_opt, vae_opt, and input_params_opt for hooking up a DiffuseHighParam node. The yaml_parameters field is a whole advanced control panel - iteration overrides, per-step VAE/sampler switching, sigma tweaks. The README calls the docs sparse; skim the YAML section anyway.

The single output is a SAMPLER - wire it into a SamplerCustom node and set your normal sigmas there.

Installing it

ComfyUI Manager, search JankDiffuseHigh and install. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/blepping/comfyui_jankdiffusehigh
cd comfyui_jankdiffusehigh
pip install -r requirements.txt

Then restart ComfyUI. Its only dependencies are pywavelets and pytorch-wavelets - Manager installs them automatically, which is refreshingly light for a node this fancy. No model files to download; TAESD mode is the sole extra if you go there.

Where people get burned

The biggest complaints are speed and VRAM. This is slow - multiple full-res iterations at high resolution - and hungry. If you see the model repeatedly unloading/reloading, switch guidance_mode to latent, and if you OOM at high res while a normal img2img fits, set ensure_model_mode: lowvram or novram in the YAML. Brownian-noise (SDE) samplers can misbehave; the author has a workaround but it's not bulletproof. SD3.5 apparently can't handle high resolutions at all, even for img2img - skip it there. And the reference image matters: this method locks onto its structure, so a mediocre first pass stays mediocre. One more gotcha: it works around a long-standing ComfyUI nondeterminism bug, which can change your seed. And if you're on SD1.5, grab the author's MSW-MSA attention from jankhidiffusion - a genuinely free speed lunch at these resolutions.