TT FLUX Models Loader (Advanced)

A stock FLUX graph needs a UNETLoader (or DiffusionModelLoader), a DualCLIPLoader for CLIP-L plus T5-XXL, a VAELoader, and then a chain of LoraLoaders - five nodes before you've even touched a text encoder. TenserTensor's FLUX loader folds all of that into one box. You pick a diffusion model, pick your two text encoders, optionally stack up to four LoRAs, and get MODEL, CLIP, and VAE out the right side.

The "Advanced" part is the useful part: apply_sampling (on by default) patches a Flux sampling-shift into the model. Flux gets washed-out, mushy output at high resolutions if you leave the schedule alone, so the node linearly interpolates a shift between base_sampling_shift (0.5, used down near 256px) and max_sampling_shift (1.15, at big resolutions) based on sampling_width/sampling_height. It's the same trick the ModelSamplingFluxAdvanced patch does, applied for you. If you're generating at 1024x1024 or above, leave it on.

The other headline input is unet_dtype. Flux is a 12B stack and most of us can't hold fp16 in memory, which is why the community runs it on fp8 or GGUF as the norm. Options are default, fp8_e4m3fn, fp8_e4m3fn_fast, and fp8_e5m2 - the first fp8 keeps compute speed, the _fast variant adds fp8 optimizations, and e5m2 is the rougher of the two. For the text side, clip_l and t5xxl pull from your text_encoders folder, and you can push clip_device to CPU if the T5 is eating your VRAM.

The LoRA slots (lora_name_1–4 with strength_1–4) are genuinely useful and have a small gotcha: strengths go from -10 to +10, so you can do negative LoRA steering, not just 0–1 blending. The pack caches loaded LoRA weights on the node instance, which means rapid re-runs are faster - but it also means a stale cache can persist if you change files on disk and don't reload. When in doubt, refresh the workflow.

Install is the standard one:

cd ComfyUI/custom_nodes
git clone https://github.com/tenser-tensor/ComfyUI-TenserTensor

or search "TenserTensor" in ComfyUI Manager and restart. The pack's real dependencies are gguf>=0.17.0 and kornia; the gguf dependency is why ComfyUI needs a recent version for UNET-only model loading.

Where people get burned:

• GGUF files don't show up. ComfyUI doesn't scan .gguf by default. This pack registers the diffusion_models_gguf and text_encoders_gguf folders automatically, so drop quantized weights there - they're picked up from unet_name, clip_l, and t5xxl just like safetensors.
• T5-XXL is the memory hog. The transformer might fit, then the 4.7B T5 blows your VRAM. Load the T5 as GGUF (Q4/Q5 is fine - the KB consensus is Q8 ≈ fp16 and it degrades gracefully) or set clip_device to CPU.
• The sampling shift isn't magic. It helps high-res Flux a lot, but it won't fix a bad scheduler choice. Euler + simple, 20–30 steps, is still the boring default that works.

One honest caveat: this class name is the pack's original (V1) node, now sitting in a Deprecated/ category while the author migrates everything to ComfyUI's newer API V3 ("Node"-suffixed replacements). It still loads and runs fine - but if the pack ever drops V1 support, the TT_FluxModelsLoaderAdvancedNode is the drop-in successor. TenserTensor is a one-person, Apache-2.0 project that barely surfaces in community discussion, so weigh that maintenance risk against how tidy the node makes your graph.