FLUX Batch Prompts

Everything that makes batch prompting annoying on SD is worse on FLUX. FLUX uses dual text encoders, guidance as a first-class parameter, and a conditioning format that doesn't always play nice with naive batch tricks. FluxBatchPrompts is the Endless pack's answer: same one-prompt-per-line, same simultaneous batch, but with FLUX's guidance scale handled for you so the output slots straight into a standard KSampler.

If you're generating FLUX and want to test 6-8 prompt variations in a single pass instead of queueing them, this is the node. The author's own use case is exactly this - batch prompting came from being tired of ComfyUI running different prompts one at a time, wasting the big VRAM buffer you've paid for.

How it works

The source shows a two-pronged strategy. It first tries true batch encoding: join all prompts into one multi-line string, tokenize the whole thing at once, and check that the resulting conditioning tensor's batch dimension actually equals your prompt count. If the CLIP supports it, you get one tensor with shape [N, ...] and a guidance value stuffed into the conditioning dict (guidance and guidance_scale keys, both set from the node's guidance input) - exactly what FLUX's sampler reads.

If that unified path fails (separate CLIP-L / T5 encoders, or a CLIP that can't batch-tokenize), it falls back to encoding each prompt individually and stacking. Either way you get the same CONDITIONING out; the node just picks the path that works on your setup.

The inputs that matter

• prompts - multiline, one prompt per line. The list is the point.
• clip - the FLUX CLIP encoder (the one from your model loader). It handles dual encoders internally.
• guidance - default 3.5, range 0-100. This is FLUX's guidance, not CFG. Most FLUX workflows live around 3.5; this is what the README says the built-in guidance scale is for.
• max_batch_size (0-64) - 0 means "match my line count." Lower truncates, higher cycles your prompts to pad the batch.
• print_output - console logging for each prompt.

Outputs: CONDITIONING → KSampler positive. PROMPT_LIST → pipe-separated prompt string for the Endless Image Saver filenames. PROMPT_COUNT → integer count.

Wiring it up

Standard FLUX graph: load model + dual CLIP, feed clip here, and send CONDITIONING to the KSampler. Keep the latent batch size equal to your prompt count (or let max_batch_size enforce a target). The README's example workflow shows it paired with a Dynamic Prompts wildcard node on the front end and "Run (instant)" queue mode - endless random batched prompts with zero babysitting.

Installing it

Part of Endless 🌊✨ Nodes by tusharbhutt. ComfyUI Manager → search "Endless" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/tusharbhutt/Endless-Nodes
# restart ComfyUI

No extra dependencies and no model downloads for this node - the pack declares no pip requirements beyond what ComfyUI ships.

Common issues

• Guidance vs CFG confusion - this node's guidance is FLUX guidance, not CFG scale. If you wire it into an SD-style CFG workflow expecting that number to behave like CFG, results will look off. FLUX lives around 3.5.
• Batch size mismatch - if your latent batch size doesn't equal the conditioning batch size, the sampler will error. Use max_batch_size to enforce alignment, or keep line counts and latent batch in sync manually.
• Dual-encoder edge cases - if your model uses separate CLIP-L and T5 encoders, the node falls back to per-prompt encoding. It still works; you just lose the single-shot tokenize. Nothing to fix, but expect slightly slower encode time.
• The pack itself - one hobbyist maintainer, AI-assisted code, rewritten mid-2025. The batch nodes are the pack's headline feature and the most battle-tested part of it. If the repo goes quiet, these still work - they're self-contained Python.

And one thing the README is refreshingly explicit about: the author forbids using any of these nodes to violate the Stable Diffusion or Black Forest Labs usage policies. Reasonable line, worth knowing before you build on the pack.