PhotoDoodle_Loader

If you've ever loaded FLUX.1-dev the conventional ComfyUI way - a diffusion-model loader, a VAE loader, a DualCLIPLoader for the T5, then one or two LoRA loaders - you know it's a chain of parts that each wants a different folder. PhotoDoodle_Loader collapses the whole chain into one node. You pick a model, pick a LoRA, and out the other end comes a single MODEL_PhotoDoodle object that only one other node on earth knows what to do with: PhotoDoodle_Sampler.

This is the "assemble everything" half of a deliberately small two-node pack. The pack is a ComfyUI port of PhotoDoodle, the Show Lab's Learning Artistic Image Editing from Few-Shot Pairwise Data (arXiv 2502.14397). The idea is straightforward: train on pairs of before/after photos so a FLUX.1-dev model learns a doodle effect - a halo and wings, a flame edge, a paint-splash filter - and can apply it to a real image while leaving the rest of the photo alone. The Loader is the unglamorous half: it loads FLUX, fuses the LoRAs, and wires up the pipeline. The Sampler is where your photo actually gets doodled on.

How it works

Under the hood the Loader assembles a diffusers FluxPipeline from whatever you hand it. The README lays out three input paths, in order of preference:

• Repo mode - put a Hugging Face id or a local diffusers folder in flux_repo (e.g. black-forest-labs/FLUX.1-dev). The author's first recommendation, and the one that auto-downloads.
• Single-file fp8 checkpoint - e.g. flux1-dev-fp8.safetensors (~16G) in models/diffusion_models. Per the README, "normal 12G can run" without extra offload tricks.
• Separate unet + vae + ComfyUI T5 - the least recommended: more likely to OOM, and it drags a CLIP dependency into the Sampler.

Then it loads pretrain.safetensors, fuses it, unloads, loads the effect LoRA, and turns on model CPU offload. That "pretrain + effect" dance is the whole trick: the pretrain LoRA is what turns vanilla FLUX into a PhotoDoodle model, and the effect LoRA is the specific style you asked for.

The inputs that matter

• flux_repo - empty string uses the single-file/unet path; an HF id or local path switches to repo mode. If you have 12–24GB, this is where you start.
• flux_unet / vae - dropdowns of your diffusion_models and vae folders. With a single-file FLUX checkpoint that bundles CLIP and VAE, set vae to "none". Pick a separate vae only for the unet+ae+clip path.
• pre_lora - the mandatory pretrain.safetensors. Quirk: the dropdown only lists LoRAs with "pre" in the filename, so keep that name.
• loras - the effect LoRA (sksmagiceffects, sksedgeeffect, …). Both LoRAs are effectively required; the code throws ValueError("No model selected") if either is "none".
• use_mmgp and profile_number - flip use_mmgp on and the loader hands the pipeline to mmgp's VRAM offload profiler, with a handful of preset profiles. It's the author's own OOM escape hatch.

The single output, model (MODEL_PhotoDoodle), wires straight into PhotoDoodle_Sampler's model input.

Installing it

ComfyUI Manager, search ComfyUI_PhotoDoodle; or:

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

then restart ComfyUI. requirements.txt pins diffusers==0.32.2 and pulls accelerate, transformers, peft, and bitsandbytes (that last one covers the fp8/nf4 quantized loading). Then download pretrain.safetensors plus the effect LoRAs from huggingface.co/nicolaus-huang/PhotoDoodle into models/loras, and your FLUX.1-dev checkpoint into models/diffusion_models.

Where people get burned

• OOM. Single-file fp8 handles roughly 12G; below that, repo mode plus use_mmgp is the lever. The README literally says "if OOM, try mmgp".
• Picking a VAE with a single-file model. Any non-"none" vae is read as "you're doing the unet path", which flips an internal flag and makes the Sampler demand a CLIP you weren't planning to provide.
• The pinned diffusers==0.32.2. If another custom node insists on a different diffusers version, one of them breaks. Welcome to dependency roulette.
• License. This builds on FLUX.1-dev, whose non-commercial license covers selling what you generate but not hosting the weights - fine for personal use, not for a paid service.