Models: where they live and the rule
Rule: all models live in one flat folder, /Volumes/SSD-4T-LR/AI/Models (exFAT SSD, no subfolders, no symlinks between models). Draw Things uses it as its External Model Folder. Check that folder before any download; never download a file that exists. New downloads go there.
Files (in /Volumes/SSD-4T-LR/AI/Models) |
Used for |
|---|---|
ltx_2.3_22b_distilled_1.1_q8p.ckpt, ltx_2.3_audio_video_vae_f16.ckpt, ltx_2.3_spatial_upscaler_x1.5_f16.ckpt, ltx_2.3_spatial_upscaler_x2_1.1_f16.ckpt, gemma_3_12b_it_qat_q8p.ckpt (+-tensordata) |
Video (LTX-2.3 image-to-video) — the model job 001 used |
ltx_2_19b_distilled_q6p.ckpt, ltx_2_audio_video_vae_f16.ckpt |
Older LTX-2 19B (not used) |
wan_v2.2_a14b_hne_t2v_q8p.ckpt, wan_v2.1_video_vae_f16.ckpt, umt5_xxl_encoder_q8p.ckpt (+tensordata) |
Wan 2.2 text-to-video high-noise expert (can't animate a still; unused) |
flux_2_klein_9b_i8x.ckpt, flux_1_vae_f16.ckpt, flux_2_vae_f16.ckpt, z_image_turbo_1.0_q8p.ckpt, qwen_3_8b_q8p.ckpt, qwen_3_vl_4b_instruct_q8p.ckpt (+tensordata), custom_prompt_style.json |
Image models (not used here) |
ACE-Step/ (9.6 GB) |
Music model weights (ACE-Step 1.5, MIT). tools/ACE-Step-1.5/checkpoints is a symlink to this folder |
Other local software: Ollama models (gpt-oss:120b 65 GB — never load during renders, qwen3.5:35B, qwen3-coder:30b, qwen3.8:27b) exist but were not used for this pipeline.
If image-to-video with Wan is ever wanted, the I2V experts must be added (not installed). Consolidation history (2026-09-23): 15 files were copied from Draw Things' sandbox to the SSD, verified by exact byte size, then deleted from the sandbox.
FLUX.2 klein 9B as an image editor (verified 2026-10-17): ~/.claude/bin/dt-cli-26.0928.0/draw-things-cli generate --local --no-download-missing --model flux_2_klein_9b_i8x.ckpt --prompt-file p.txt --image crop.png --strength 1.0 --width 1280 --height 1280 --seed S edits a crop in 4 steps (~13 s at 1280², ~6 s per step on a busy machine). Give it a crop (≤ 1280 px), say what to switch off and "keep everything else exactly the same"; always use --local (otherwise the CLI may use Draw Things cloud compute). Alignment with the original is good (median flow 0.2–1 px) and it can switch a lamp off convincingly, but pasting its pixels back into a finished frame never looked right (the edit repaints the surroundings, so a patch shows). Use it for stills or for ideas, not for patches inside an approved picture (Lessons S11–S13). Depth Anything V2 Small (depth_map.py, run with Tools/moss-env) gives the depth map the snow needs.