Shima Sampler

The Shima Sampler is the pack's workhorse: a full drop-in replacement for the core KSampler with one very Shima twist - it can eat the pack's BNDL bundles directly. Wire in a modelcitizen.bndl, a masterprompt.bndl, and a latentmaker.bndl, and the sampler pulls model, VAE, conditioning, and latent out of the bundles itself. No six loose wires, no intermediate DeBNDLers. It also bundles a live latent preview during sampling and a built-in two-pass latent upscale, so it's a genuinely bigger node than the KSampler it replaces - for better and for worse.

How it works

Under the hood it constructs a standard ComfyUI KSampler and runs it, so everything you know about sampling applies. The Shima layer is in the plumbing around it:

• Seed is s33d, not seed - the pack deliberately renames it so ComfyUI's automatic control_after_generate widget never silently re-rolls your seed out from under you. The randomize toggle handles random seeds explicitly. If a shima.commonparams bundle is connected with use_commonparams on, the bundle's seed overrides the widget.
• Bundle overrides - modelcitizen.bndl supplies model + VAE, latentmaker.bndl supplies the latent, masterprompt.bndl supplies positive/negative. Each overrides its individual-wire counterpart. There's also panelinputs.bndl for the pack's Panel system.
• Two-pass HiRes Fix is built in - turn on upscale_enabled and the node upscales the latent (nearest-exact/bilinear/area/bicubic/bislerp at upscale_factor) then runs a second sample pass at upscale_denoise/upscale_steps/upscale_cfg. One node does what normally takes a LatentUpscale + second KSampler + masking rig.
• Live previews - with preview_method on auto/latents, it decodes latent previews during sampling (throttled to ~every 500ms plus the final step) and shows them inline, instead of a blank node until it finishes.

It raises loud, useful errors when you feed it nothing: "No Model provided!", "No Conditioning provided!", "No Latent Image provided!" - which makes wiring mistakes obvious fast.

The inputs that matter

• model, positive, negative, latent - the classic four, or the equivalent .bndl inputs.
• s33d / randomize - seed control.
• steps, cfg, sampler_name, scheduler, denoise - the standard knobs; use_samplercommons lets a Shima Sampler Commons bundle override these.
• upscale_enabled / upscale_factor / upscale_denoise / upscale_steps / upscale_cfg - the built-in HiRes Fix.
• vae_decode - decode to an image (needs a VAE from the model, the vae input, or the model citizen bundle).
• use_commonparams + shima.commonparams - pull seed/model-type from Shima.Commons.

The outputs

• latent - the sampled latent.
• image - the decoded image (a 64×64 placeholder if vae_decode is off).
• s33d - the seed actually used, which matters when randomization or Commons picked it.
• shimasampler.bndl - the result envelope (image + latent + used seed) for downstream Shima nodes.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/KDB-USJP/shima_wf.git Shima
pip install -r requirements.txt

Restart, then Shima/Sampling → Shima Sampler. ComfyUI Manager: search "Shima". On first boot the installer auto-clones ComfyUI-Impact-Pack and cg-use-everywhere if missing (the Use Everywhere dependency for island auto-linking) - expected.

Common issues

The big one is stale bundles. The sampler reads the bundle at execution time, so if you build a masterprompt.bndl before a LoRA or style node and never rebuild it, the sampler samples with the old conditioning and nothing looks broken - the snapshot rule from the plumbing essay in full effect. Second: use_commonparams defaults on, so a connected Commons bundle silently overrides your carefully typed seed and settings; if the sampler ignores your widget values, check what the Commons node is broadcasting. (use_samplercommons is off by default - Sampler Commons overrides only kick in when you switch it on.) Third, if image comes out as a 64×64 black box, vae_decode is on but no VAE reached the node - the model citizen bundle is the usual culprit. It's a new, niche pack with thin community coverage, so when behavior surprises you, the sample logic in nodes/sampler.py is clear enough to read end-to-end.