Text-to-Image
Diffusers
Safetensors
MiniMax H3
VideoX Fun
controlnet
controlnet-union
image-to-image
ml-intern
Instructions to use akhaliq/MiniMax-H3-Image-Fun-Controlnet-Union-2.0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use akhaliq/MiniMax-H3-Image-Fun-Controlnet-Union-2.0 with Diffusers:
pip install -U diffusers transformers accelerate
from diffusers import ControlNetModel, StableDiffusionControlNetPipeline controlnet = ControlNetModel.from_pretrained("akhaliq/MiniMax-H3-Image-Fun-Controlnet-Union-2.0") pipe = StableDiffusionControlNetPipeline.from_pretrained( "MiniMaxAI/MiniMax-H3-Image", controlnet=controlnet ) - VideoX Fun
How to use akhaliq/MiniMax-H3-Image-Fun-Controlnet-Union-2.0 with VideoX Fun:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
- DiffusionBee
merge/: hint spatial-structure probe (offline control-branch replay)
Browse files- merge/diag_hints.py +152 -0
merge/diag_hints.py
ADDED
|
@@ -0,0 +1,152 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Probe: do the control hints carry the control image's spatial structure?
|
| 2 |
+
|
| 3 |
+
Captures real intermediates (packed seq, temb, adaln_indices, rotary_emb,
|
| 4 |
+
image_indices) from one denoise forward, then replays the control branch
|
| 5 |
+
offline with circle-left vs circle-right control rows and reports per-row
|
| 6 |
+
hint norms as 32x32 maps. Decides between:
|
| 7 |
+
A. control rows/hints are spatially flat -> wiring/prep bug (fixable)
|
| 8 |
+
B. hints carry structure but DiT ignores -> training-trajectory mismatch
|
| 9 |
+
"""
|
| 10 |
+
import os
|
| 11 |
+
|
| 12 |
+
os.environ.setdefault("HF_XET_HIGH_PERFORMANCE", "1")
|
| 13 |
+
|
| 14 |
+
import cv2
|
| 15 |
+
import numpy as np
|
| 16 |
+
import torch
|
| 17 |
+
from PIL import Image
|
| 18 |
+
from huggingface_hub import hf_hub_download
|
| 19 |
+
from safetensors import safe_open
|
| 20 |
+
from gradio_client import Client
|
| 21 |
+
|
| 22 |
+
from validate_controlnet import (
|
| 23 |
+
attach_control_branch,
|
| 24 |
+
wrap_forward_with_control,
|
| 25 |
+
build_control_rows,
|
| 26 |
+
log,
|
| 27 |
+
ENCODER_SPACE,
|
| 28 |
+
BASE,
|
| 29 |
+
CONTROLNET_REPO,
|
| 30 |
+
CONTROLNET_FILE,
|
| 31 |
+
SEED,
|
| 32 |
+
)
|
| 33 |
+
|
| 34 |
+
PROMPT = "A rough weathered concrete wall texture filling the entire frame, soft even light, photorealistic."
|
| 35 |
+
OUT = "/work/diag2"
|
| 36 |
+
os.makedirs(OUT, exist_ok=True)
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def circle_control(cx):
|
| 40 |
+
img = np.zeros((1024, 1024), np.uint8)
|
| 41 |
+
cv2.circle(img, (cx, 512), 200, 255, 8)
|
| 42 |
+
return np.repeat(img[..., None], 3, axis=-1)
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
def norm_map(rows_2d, tag):
|
| 46 |
+
"""rows_2d: (1024, C) -> 32x32 L2-norm map + left/right stats + heatmap png."""
|
| 47 |
+
norms = rows_2d.float().norm(dim=-1).reshape(32, 32).cpu().numpy()
|
| 48 |
+
l, r = norms[:, :16].mean(), norms[:, 16:].mean()
|
| 49 |
+
log(f"{tag}: row-norm left={l:.4f} right={r:.4f} ratio L/R={l/max(r,1e-9):.3f} "
|
| 50 |
+
f"min={norms.min():.4f} max={norms.max():.4f}")
|
| 51 |
+
heat = ((norms - norms.min()) / max(np.ptp(norms), 1e-9) * 255).astype(np.uint8)
|
| 52 |
+
Image.fromarray(heat).resize((256, 256), Image.NEAREST).save(f"{OUT}/{tag}_heatmap.png")
|
| 53 |
+
return norms
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
def main():
|
| 57 |
+
log("loading pipeline (cached)")
|
| 58 |
+
ckpt = hf_hub_download(CONTROLNET_REPO, CONTROLNET_FILE)
|
| 59 |
+
from diffusers import ModularPipeline
|
| 60 |
+
|
| 61 |
+
blocks = ModularPipeline.from_pretrained(BASE).blocks
|
| 62 |
+
blocks.sub_blocks.pop("text_encoder")
|
| 63 |
+
pipe = blocks.init_pipeline(BASE)
|
| 64 |
+
pipe.load_components(dtype=torch.bfloat16)
|
| 65 |
+
model = pipe.transformer
|
| 66 |
+
attach_control_branch(model, ckpt)
|
| 67 |
+
wrap_forward_with_control(model)
|
| 68 |
+
pipe.to("cuda")
|
| 69 |
+
|
| 70 |
+
# capture: forward kwargs (image_indices...) + block-0 inputs (packed, temb, adaln, rope)
|
| 71 |
+
cap = {}
|
| 72 |
+
orig_forward = model.forward
|
| 73 |
+
def rec(hidden_states, encoder_hidden_states, timestep, timestep_indices,
|
| 74 |
+
token_tags, position_ids, image_indices, text_indices, pdd_step_index,
|
| 75 |
+
return_dict=True, control_rows=None, control_context_scale=1.0):
|
| 76 |
+
cap.setdefault("image_indices", image_indices)
|
| 77 |
+
cap.setdefault("text_indices", text_indices)
|
| 78 |
+
cap.setdefault("position_ids", position_ids)
|
| 79 |
+
return orig_forward(hidden_states, encoder_hidden_states, timestep,
|
| 80 |
+
timestep_indices, token_tags, position_ids, image_indices,
|
| 81 |
+
text_indices, pdd_step_index, return_dict=return_dict,
|
| 82 |
+
control_rows=control_rows, control_context_scale=control_context_scale)
|
| 83 |
+
model.forward = rec
|
| 84 |
+
|
| 85 |
+
block0 = {}
|
| 86 |
+
def pre_hook(m, args):
|
| 87 |
+
if "packed" not in block0:
|
| 88 |
+
block0["packed"], block0["temb"], block0["adaln"], block0["rope"] = (
|
| 89 |
+
args[0].detach(), args[1].detach(), args[2].detach(),
|
| 90 |
+
(args[3][0].detach(), args[3][1].detach()) if isinstance(args[3], tuple) else args[3].detach(),
|
| 91 |
+
)
|
| 92 |
+
h = model.transformer_blocks[0].register_forward_pre_hook(pre_hook)
|
| 93 |
+
|
| 94 |
+
client = Client(ENCODER_SPACE, token=os.environ.get("HF_TOKEN"), httpx_kwargs={"timeout": 600})
|
| 95 |
+
embeds_path = client.predict(PROMPT, [], api_name="/encode")
|
| 96 |
+
if isinstance(embeds_path, (tuple, list)):
|
| 97 |
+
embeds_path = embeds_path[0]
|
| 98 |
+
if isinstance(embeds_path, dict):
|
| 99 |
+
embeds_path = embeds_path["path"]
|
| 100 |
+
with safe_open(embeds_path, framework="pt") as f:
|
| 101 |
+
prompt_embeds = f.get_tensor("prompt_embeds")
|
| 102 |
+
text_token_tags = f.get_tensor("text_token_tags")
|
| 103 |
+
|
| 104 |
+
ratio = getattr(pipe, "vae_spatial_compression_ratio", 16)
|
| 105 |
+
rows_left = build_control_rows(pipe, circle_control(256), 1024 // ratio, 1024 // ratio)
|
| 106 |
+
|
| 107 |
+
log("one generation to capture intermediates")
|
| 108 |
+
pipe(
|
| 109 |
+
prompt_embeds=prompt_embeds.to("cuda", torch.bfloat16),
|
| 110 |
+
text_token_tags=text_token_tags,
|
| 111 |
+
width=1024, height=1024, min_generation_area=1024,
|
| 112 |
+
generator=torch.Generator("cpu").manual_seed(SEED),
|
| 113 |
+
output="images",
|
| 114 |
+
denoiser_input_fields={"control_rows": rows_left, "control_context_scale": 1.0},
|
| 115 |
+
)
|
| 116 |
+
h.remove()
|
| 117 |
+
model.forward = orig_forward
|
| 118 |
+
assert "packed" in block0 and "image_indices" in cap, (list(block0), list(cap))
|
| 119 |
+
log(f"captured: packed {tuple(block0['packed'].shape)}, image rows {cap['image_indices'].shape[0]}")
|
| 120 |
+
|
| 121 |
+
rows_right = build_control_rows(pipe, circle_control(768), 1024 // ratio, 1024 // ratio)
|
| 122 |
+
|
| 123 |
+
with torch.no_grad():
|
| 124 |
+
for tag, rows in (("left", rows_left), ("right", rows_right)):
|
| 125 |
+
# stage 1: control_proj_in output
|
| 126 |
+
ce = model.control_proj_in(rows.to(next(model.control_proj_in.parameters()).dtype))
|
| 127 |
+
norm_map(ce[0], f"projin_{tag}")
|
| 128 |
+
# stage 2: replay control blocks on the captured intermediates
|
| 129 |
+
c = block0["packed"].index_copy(1, cap["image_indices"], ce.to(block0["packed"].dtype))
|
| 130 |
+
hints = []
|
| 131 |
+
for n, blk in enumerate(model.control_blocks):
|
| 132 |
+
if n == 0:
|
| 133 |
+
c = blk.before_proj(c) + block0["packed"]
|
| 134 |
+
c = blk(c, block0["temb"], block0["adaln"], block0["rope"])
|
| 135 |
+
hints.append(blk.after_proj(c))
|
| 136 |
+
img_rows = cap["image_indices"]
|
| 137 |
+
norm_map(hints[0][0, img_rows], f"hint0_{tag}")
|
| 138 |
+
norm_map(hints[9][0, img_rows], f"hint9_{tag}")
|
| 139 |
+
del hints
|
| 140 |
+
torch.cuda.empty_cache()
|
| 141 |
+
|
| 142 |
+
from huggingface_hub import HfApi
|
| 143 |
+
HfApi().upload_folder(
|
| 144 |
+
folder_path=OUT, path_in_repo="assets/diag2",
|
| 145 |
+
repo_id="akhaliq/MiniMax-H3-Image-Fun-Controlnet-Union-2.0", repo_type="model",
|
| 146 |
+
commit_message="Hint spatial-structure probe (control branch replay)",
|
| 147 |
+
)
|
| 148 |
+
log("DONE")
|
| 149 |
+
|
| 150 |
+
|
| 151 |
+
if __name__ == "__main__":
|
| 152 |
+
main()
|