FluxFill Pipeline Loader [ComfyUI-Flux-TryOff]

This is the node that makes ComfyUI-Flux-TryOff feel magic after the original catvton-flux felt like a part-time job. Where the legacy FluxFill Model Loader demands you already own FLUX.1-dev, this one assembles the entire pipeline - text encoders, VAE, scheduler, transformer - and downloads whatever's missing on first run. It's the recommended path in every current workflow.

What it builds

TryOffFluxFillPipelineNode (registered as "FluxFill Pipeline Loader") constructs a diffusers FluxFillPipeline piece by piece:

• Text encoders - CLIP-L (openai/clip-vit-large-patch14) and the big T5 (XLabs-AI/xflux_text_encoders) that Flux needs to understand full-sentence prompts. Cached into ComfyUI/models/text_encoders so you download them once.
• VAE - madebyollin/taef1, the tiny TAESD1 autoencoder, cached into ComfyUI/models/vae. TAESD1 is the small, fast, low-VRAM decoder - the reason previews render quickly but can look a bit soft. It's a deliberate trade for fitting on 16GB cards.
• Scheduler - a FlowMatchEulerDiscreteScheduler, which is the right scheduler for Flux. Don't go hunting for Karras; flow-matching Flux wants plain Euler.
• transformer (required input) - the MODEL from the TryOff Model Loader. That's the CatVTON fine-tune doing the real work; everything above is just its support staff.

Like the legacy loader, it calls enable_model_cpu_offload() after assembly, which is how a 12B transformer plus T5 coexist with a 4080.

The inputs

• transformer - wire it from the TryOff Model Loader. Non-negotiable.
• device - cuda, obviously, unless you hate yourself.
• transformers_config / diffusers_config (both optional) - the two outputs of the TryOff Quantizer. Wire both for 8-bit/4-bit. They go in as a matched pair; that's exactly why the quantizer outputs two configs.

The output is a single MODEL - but this one isn't a transformer, it's the whole pipeline object. It wires straight into the pipe input of Run TryOn Inference or Run TryOff Inference. Yes, the type is MODEL; no, it doesn't work in a regular KSampler. This pipe is only for these run nodes.

Download behavior (the part people trip on)

Everything pulls from Hugging Face automatically, no token needed - the transformer into models/checkpoints, encoders into models/text_encoders, VAE into models/vae. That's two encoders plus a VAE plus whatever model you chose, so the first run downloads a serious chunk of gigabytes. Let it finish before assuming something's broken; the node shows nothing until the pipeline object exists. And "downloaded to ComfyUI's model folders" means you'll see files land in text_encoders and vae that you didn't put there - that's this node, working as intended.

Install

ComfyUI Manager → search "ComfyUI-Flux-TryOff", or:

cd ComfyUI/custom_nodes
git clone https://github.com/asutermo/ComfyUI-Flux-TryOff

Restart, then diffusers>=0.32.2 and friends install. If you quantize, bitsandbytes needs to be present even though it's not in the pack's requirements.txt. First run is slow because of downloads; every run after is slow because the author says so - tested on a 4080, wants a 4090.