Meld Unified Flux Loader

A normal Flux workflow makes you wire four or five nodes just to get a model, a text encoder, and a VAE into the graph. Meld Unified Flux Loader collapses that into one box. Pick your UNet, pick your two CLIPs, pick your VAE, type the prompt - it hands you MODEL, CLIP, VAE, the prompt string, and a bundled base_settings dict, all in one go.

It's part of ComfyUI-Meld, the pack built around the Meld Image Manager. The pitch is "turn your generations into an asset database," and this node is the load side of that loop: one tidy node that defines both the model and every generation parameter, so settings stay in one place instead of being scattered across the canvas.

What it actually does

The node's real trick is extension-sniffing. The model_name dropdown merges everything ComfyUI knows about from models/unet, models/checkpoints, and (if the ComfyUI-GGUF pack is installed) the GGUF loader's file list. Pick something ending in .gguf and the node silently swaps to the GGUF UNet loader; pick a .safetensors or .ckpt and it routes to the standard loader. Same deal for CLIP: give it a .gguf clip and it calls DualCLIPLoaderGGUF instead of the regular one. You don't pick a loader - the file extension does. That's genuinely handy if you run Q4/Q8 Flux to fit VRAM, since the quantized and full-precision paths behave differently.

Two Flux-specific opinions are baked in. CFG is fixed to 1.0 and the negative prompt is hardcoded to empty, which matches how Flux's guidance-distilled models actually want to be run. And width/height are rounded down to a multiple of 16 internally to avoid the classic Flux padding errors - type 1032 and it quietly makes it 1024.

The inputs that matter

• model_name - your Flux UNet. Pick the .gguf file if that's what you run.
• weight_dtype - default or one of the fp8 options (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2). Matters only for non-GGUF UNets.
• clip_name1 / clip_name2 / clip_type - normally t5xxl_fp16.safetensors plus clip_l.safetensors with clip_type set to flux.
• guidance - Flux guidance scale, default 3.5. This is the field the standard Flux workflow hides in a custom node; here it's just a number.
• seed, steps, sampler_name, scheduler, width, height - the usual sampling settings.

Outputs are MODEL, CLIP, VAE, positive, and base_settings. That last one is a dict you can feed straight into Meld Settings Unpacker to pull out seed/steps/guidance/sampler/scheduler/resolution as individual wires.

Installing it

This ships in ComfyUI-Meld, so install the pack once:

• ComfyUI Manager → search "Meld" → install, or
• comfy node install HappyOnigiri/ComfyUI-Meld, or
• manually: cd ComfyUI/custom_nodes && git clone https://github.com/HappyOnigiri/ComfyUI-Meld.git && cd ComfyUI-Meld && pip install -r requirements.txt then restart.

You'll also need the Flux model files themselves (the UNet in models/unet, t5xxl + clip_l in models/clip, the AE in models/vae), and - if you want the GGUF switching to actually engage - the ComfyUI-GGUF custom node pack installed.

Gotchas

The GGUF auto-switch depends on ComfyUI-GGUF being present; without it the node falls back to the standard loader and a .gguf pick will likely fail, so install both or neither. And remember the developer is Japanese and answers GitHub issues via machine translation - simple English plus a screenshot or log goes a long way. One light aside: if you're on the fence about quantization, Q8 is basically fp16 at half the size - the auto-switch makes it painless to test both.