🔀 Batch Wildcard Upscale Sampler

This is the biggest, most opinionated node in the whole pack, and it solves a real limitation of how ComfyUI batches normally work. A standard batch shares one resolved prompt across every image in it - if your prompt has a wildcard in it, ComfyUI resolves it once and every image in the batch gets the same result. This node resolves wildcards, and any inline LoRA tags, separately for every single image, then runs them one at a time internally and stitches the results back into one batch-shaped output at the end. You get true per-image prompt variation without giving up the convenience of a single batch downstream. The README credits the underlying approach to ChronoKnight's original code, adapted here with LoRA-tag loading and an optional hi-res pass bolted on.

The core trade-off is stated honestly in the tooltip rather than hidden: this is not a true parallel sampler batch. Processing sequentially inside one node is still faster in practice than the obvious alternative - separate nodes per image, manually wired - but it's not the raw-throughput win a native batch gives you. What you're buying is variation, not speed.

The wildcard syntax is the same one A1111 popularized years ago and that a lot of people who came to ComfyUI after 2023 never actually learned: {a|b|c} picks randomly between options, {5::black|green|red} weights the choice (black five times more likely), {2$$a|b|c|d} pulls a specific count of items out instead of one, and __filename__ pulls a random line out of a wildcard text file. <lora:name:strength> tags work inline in the prompt too - fuzzy-matched by filename against your installed LoRAs, auto-stripped before the text reaches CLIP, the same convention as this pack's standalone LoRA Loader Prompt Tags node.

The fields a beginner actually touches: text and negative are your prompts, wildcards and all. seed is the base - image 1 uses seed+0, image 2 uses seed+1, and so on, so the same seed reliably reproduces the same sequence of resolved prompts. batch_size, width, height, steps, cfg, sampler_name, scheduler, and denoise are the standard KSampler set you already know. model and clip are optional - leave them disconnected and you can preview resolved prompts for free, with no sampling and no GPU spend, which is a genuinely useful way to test whether your wildcard combos are producing what you expect before committing compute to them.

The optional hi-res pass is its own small subsystem: flip upscale on and the first-pass latent gets decoded, upscaled, and sampled again at upscale_rate times the original size, with its own upscale_denoise, upscale_steps, upscale_cfg, sampler, and scheduler (each can independently reuse the first pass's setting via a "(same as first pass)" option). This pass needs a vae connected - without one, the tooltip is explicit that the whole upscale step is silently skipped, which is the single easiest way to think you've enabled hi-res and not gotten it. If you connect an upscale_model (a Load Upscale Model node feeding in something like a 4x ESRGAN), it runs the AI upscale first and then downsamples to the exact size your upscale_rate implies - so a 4x model with upscale_rate at 2.0 still lands you a clean 2x result rather than a blown-out 4x one.

strip_prompt_weights deserves a specific callout: it strips (word:1.3)-style attention syntax before encoding, and you want it on for any modern LLM-encoded model family (Flux, SD3-lineage and similar) - those text encoders read prompts as instructions, not as CLIP token soup, so parenthesis-and-colon weighting isn't ignored, it's passed through as literal punctuation and actively hurts your result. Leave it off for CLIP-era checkpoints where weighting still works as intended.

Outputs: model/clip/vae pass through patched with whichever LoRA the last batch item resolved - worth knowing if you're batching four images with four different LoRA tags, because only the last one's patch survives on this output. positive/negative conditioning are likewise from the last item. latent is the actual combined batch you care about, and prompt gives you the resolved text for every item as a list - genuinely useful for auditing which wildcard combination produced which image after the fact.

Where this breaks in practice: a mistyped LoRA name in a tag fails the same way it does everywhere else in the pack - silently, via fuzzy matching that just doesn't find a close enough hit. Forgetting the vae for the upscale pass is the other big one. And running weighted prompt syntax against an LLM-encoded model without strip_prompt_weights on is a subtle one that produces plausible-but-wrong output rather than an obvious error.