LC Model Optimizer ⚡

The hardest part of ComfyUI in 2026 isn't prompting. It's the file menu. Anima wants qwen_3_06b_base, Krea 2 wants a Qwen3-VL 4B, Klein 9B wants a Qwen3 8B, and the diffusion model itself could be bf16, fp8, nvfp4, INT8 ConvRot or a GGUF at Q8/Q6/Q4/Q2 - with the text encoder as a second VRAM budget you have to fit alongside it. Get it wrong and you either OOM or quietly run a mismatched encoder.

LC Model Optimizer is a loader that answers the "which file, and does it fit?" question for you. It replaces your model, text encoder and VAE loaders with one node, and it picks based on a report actually measured on your card.

How it decides

The node leans on the pack's System & Model Optimization Report (Settings → LC123 Settings ⚙️ → Optimization, or right-click the empty canvas). That report tests your hardware - GPU, VRAM, RAM, page file, drive speed, PCIe link - and actually runs the speed-up stack rather than looking it up: Sage, Triton, Sol-Attn, Comfy Kitchen, fp8/int8/nvfp4, llama-cpp. It takes about ten seconds.

You pick two things:

• base_model - Krea 2 Turbo, Krea 2 Raw, Qwen-Image 2.1, Z-Image Turbo, Flux.2 Klein base 9B, Ideogram 4, Anima, SDXL 1.0, Illustrious XL, Pony Diffusion V6 XL, or Custom. The SDXL family loads as one all-in-one checkpoint, so its text encoder and VAE come from the checkpoint.
• goal - Quality (closest to bf16), Optimal (fast and near identical), or Fast (fastest that still looks right). This is the same Quality/Optimal/Fast ladder the report uses when it tells you what file type to get.

Every file widget then defaults to ★ Recommended, which names the exact file the report picked for your machine. No report yet? It falls back to Comfy's own default files and shows a 💡 telling you to run it. The pickers browse your model folders with search, offer ⬇ Download entries for recommended files you don't have yet (downloaded on the first run), list what's already on disk for that model - including versioned finetunes like Krealism and Animosity - and then show all your folders, so third-party merges work too.

summary is a STRING output describing what it loaded, which is worth wiring to a Show Text node when you're debugging a new model.

Speed-ups, and why auto is usually right

The interesting half is the switch panel at the bottom:

• attention - auto picks the fastest backend that passed the report's test (pytorch, sage, sage3, flash, xformers, comfy_kitchen_int8). Note the word "tested": a Sage build that runs and outputs grey noise gets flagged, not trusted. Given how many machines Sage Attention silently corrupts on - Z-Image Base the loudest case - handing that decision to something that measures beats copying a config off Reddit.
• fp16_accumulation - faster fp16 math with a small quality cost. auto = Fast goal only.
• step_cache - EasyCache, which skips steps that barely change. auto = Fast goal, 20+ steps.
• torch_compile - off by default. Needs Triton; first run is slow.
• qwen21_cache - Qwen-Image 2.1 only: where the KV cache lives and how it's stored. int8 halves it at roughly bf16 accuracy; int4 quarters it at about double the per-step error.

Set speed_ups to Manual and you choose each one yourself.

Install

The pack has no requirements.txt - it runs on the torch and numpy ComfyUI already ships.

cd ComfyUI/custom_nodes
git clone https://github.com/lonecatone23/ComfyUI_LC123_nodes.git

Restart, or install via ComfyUI Manager by searching ComfyUI LC123 Nodes. Optional companions: ComfyUI-GGUF if you want quantized files (the node falls back to the calcuis gguf pack if a GGUF is refused), and Triton if you want torch_compile.

Where people get burned

Downloads are the fragile bit. The ⬇ entries pull from Hugging Face on first queue. Gated repos - MiniMax H3 and friends - require you to accept the licence on the repo page and be logged in locally, otherwise the run dies with a "this repo is gated" error mid-queue. Expect the first run after picking a download to look like a hang; it's a multi-gigabyte fetch with a progress line in the console.

A mismatched encoder won't stop you. If you hand it a text encoder that doesn't belong to the base model - Qwen3 8B on Krea 2, say - it reads the file header and shows a ⚠️ hint. It still runs it. That's deliberate: finetunes and odd merges are legitimate, and the check costs one header read rather than a model load. But the warning is information, not decoration.

Custom means custom. On base_model: Custom, recommendations, estimates and file checks all say unsupported, and clip_type becomes your job - that's how the encoder gets loaded, the same value as CLIPLoader's type (chroma, for instance, on Chroma). auto handles it when the encoder's shape matches a known model.