zjwfufu/SMD-Prompt-Sample-Resources

SMD Fixed Sampling Resources

Part of SMD-data.

Fixed inputs, reference examples and precomputed conditions for SMD's prompt sample callback. These support visual comparisons during training; they are not an evaluation set, model weights, or a general training dataset. The separate batch sample callback reads the current training batch.

Original bytes and paths are preserved under assets/prompt_samples/, assets/examples/h3/ and assets/hdri/. About 1.76 GiB total; download only the families you use.

Family (--model)	Preview cases	Runtime MiB	With rebuild inputs MiB
flux1	12	96.1	96.1
qwen_image	12	3.1	3.1
qwen_image_edit	8	78.3	92.6
wan13	8	64.0	64.0
wan14	8	64.0	64.0
hunyuan3d_shape	8	63.1	75.1
trellis2_texture	8	352.0	561.3
sd35	12	62.6	62.6
flux2_4b	12	180.0	180.0
lens	12	18.6	18.6
wan13_causal	8	64.0	64.0
trellis2_shape	8	338.8	548.1
h3_t2va	2	11.0	11.0
h3_ref2va	2	167.3	167.3

SD3.5's six experiments share sd35; FLUX.2 and Lens teacher variants share their respective student caches. H3 4/8-step variants share one cache per mode. TRELLIS Shape/Texture share raw reconstruction inputs. Experiment aliases and exact file hashes are in resource_manifest.json.

Download

Install huggingface_hub, authenticate with hf auth login, and set SMD_ROOT to your checkout. This resolves one commit, then pins every download to it; retain the printed revision for future runs.

export SMD_ROOT=/path/to/SMD
python - <<'PYCODE'
import os, subprocess, sys
from huggingface_hub import HfApi, hf_hub_download
repo = "zjwfufu/SMD-Prompt-Sample-Resources"
revision = HfApi().dataset_info(repo).sha
print("Resource revision:", revision)
tool = hf_hub_download(repo, "download.py", repo_type="dataset", revision=revision)
subprocess.run([sys.executable, tool, "--revision", revision,
                "--root", os.environ["SMD_ROOT"], "--model", "sd35"], check=True)
PYCODE

Replace sd35 with a family or experiment name; multiple names are accepted. Use --scope rebuild to include original images/features, or --list to inspect the subset. Downloads verify SHA-256 and refuse to overwrite differing local files. Run this before loading a training configuration.

Runtime includes H3 example dependencies needed by its default training loader. Ref2VA keeps release/, qwen_conditions/ and reference_av/ together. TRELLIS Texture also includes forest.exr for preview lighting.

Caches are tied to the recorded preprocessing code, model settings and input signatures. Downloading them does not make incompatible code compatible. The existing SMD read_cache checks must remain enabled.

SD3.5 reward teachers and their sources are maintained in SMD-SD3.5-Medium. Existing --teachers download commands remain supported.

Original notices are retained. The procedural HDRI is CC0; no blanket license is assigned to all example materials or encoded conditions. Their redistribution terms still require release review.