Krea2 SVDQuant Quantize All-in-One

The short version

The modular Krea 2 setup is three files scattered across three folders: the DiT in diffusion_models/, the Qwen3-VL text encoder in text_encoders/, and the VAE in vae/ - about 15.5 GB in total. This node bakes all three into one combined checkpoint (~12 GB) in models/checkpoints/, and it quantizes the text encoder while it's at it. Result: one file, one loader, one-click generation. If you've used the Krea2 SVDQuant Quantize node, this is the sequel that also handles the other two components.

Run workflows/krea2_quantize_all_in_one.json in ComfyUI to do it in the graph, or tools/build_all_in_one.py from a terminal - same underlying logic. The output loads with the Krea2 SVDQuant Checkpoint Loader, which hands you model, clip, and VAE from one node.

The inputs that matter

• source_dit - the Krea 2 DiT. Either a BF16 source (~24 GB, which gets quantized first) or an already-quantized SVDQuant/W4A4/INT8 file. The low-rank branch inputs (rank, refine_iters, act_stats, seed) only apply if you feed it BF16.
• text_encoder - the Qwen3-VL 4B BF16 safetensors (~8.88 GB). Its 252 linear projections get quantized to te_format, shrinking it to ~3.2 GB while the vision tower and embeddings stay untouched.
• vae - the Krea 2 VAE (~0.51 GB, kept unquantized).
• format - the diffusion model format, same choices as the plain Quantize node: svdq (default, 4-bit + low-rank branch), w4a4, int8, fp8.
• te_format - text encoder format: w4a4 (convrot_w4a4, ~3.2 GB) or int8 (~5.2 GB).

The rest - variant (turbo/base/unknown, mostly affects the output filename), output_name, overwrite - behave the same as the plain Quantize node. output_name defaults to something derived from variant and format, so you can usually leave it empty.

Output is a single summary string: size, tensor counts, and loader instructions for the baked file.

How to install

cd ComfyUI/custom_nodes && git clone https://github.com/alperktt/Krea-2-SVDQuant-ComfyUI krea-2-svdquant

No Python dependencies - the tooling in this pack is pure Python and builds on ComfyUI's own comfy_kitchen backend. From the terminal, the equivalent is:

cd ComfyUI/custom_nodes/krea-2-svdquant
python tools/build_all_in_one.py \
    --dit  models/diffusion_models/Krea2-Turbo-SVDQuant-W4A4-rank64.safetensors \
    --text-encoder models/text_encoders/qwen3vl_4b_bf16.safetensors \
    --vae  models/vae/Krea2-HD-vae.safetensors \
    --te-format w4a4 --variant turbo

The writer streams, so it never holds the combined 12+ GB in memory at once.

Gotchas

• Budget the time. Quantizing the DiT from BF16 means the same 54s–6min conversion the plain Quantize node takes, and the text encoder pass on top. This node blocks the queue and takes the GPU while it runs.
• The cu130 gate applies to everything here. If your torch was built against CUDA < 13, ComfyUI disables the comfy_kitchen CUDA backend and the baked checkpoint will run on a slow fallback. Check python -c "import torch; print(torch.version.cuda)" before you burn a build on it.
• Re-saving a baked checkpoint logs "left over keys in diffusion model" in the stock loader. Expected: the svdq_l1/svdq_l2 keys are what let the file round-trip into this pack's loader, and stock UNETLoader just doesn't know them. Harmless.