redux_experiment_final/scripts/run_sweeps.py

#!/usr/bin/env python3
"""Phase A calibration sweeps on inputs_v2, seed 777. Outputs under sweeps/ prefixes."""
import json, os, shutil, sys, time, urllib.request

HOST = "http://127.0.0.1:7865"
SEED = 777
W, H = 896, 1152
HW, HH = 1344, 1728
OUT_ROOT = "/workspace/outputs_v2/sweeps"
COMFY_OUT = "/workspace/ComfyUI/output"

PAIRS = [("r1.jpg", "t2.jpg"), ("r2.jpg", "t5.jpg"), ("r4.jpg", "t3.jpg")]  # variety subset

def post(prompt):
    req = urllib.request.Request(f"{HOST}/prompt",
        data=json.dumps({"prompt": prompt}).encode(), headers={"Content-Type": "application/json"})
    with urllib.request.urlopen(req) as r:
        d = json.loads(r.read())
    if "prompt_id" not in d:
        raise RuntimeError(d)
    return d["prompt_id"]

def core(p, style_img, comp_img, guidance=3.5, redux=0.5):
    p["u"] = {"class_type": "UNETLoader", "inputs": {"unet_name": "flux1-dev.safetensors", "weight_dtype": "default"}}
    p["c"] = {"class_type": "DualCLIPLoader", "inputs": {"clip_name1": "t5xxl_fp16.safetensors",
        "clip_name2": "clip_l.safetensors", "type": "flux", "device": "default"}}
    p["v"] = {"class_type": "VAELoader", "inputs": {"vae_name": "ae.safetensors"}}
    p["txt"] = {"class_type": "CLIPTextEncode", "inputs": {"clip": ["c", 0], "text": ""}}
    p["guid"] = {"class_type": "FluxGuidance", "inputs": {"conditioning": ["txt", 0], "guidance": guidance}}
    p["neg"] = {"class_type": "ConditioningZeroOut", "inputs": {"conditioning": ["guid", 0]}}
    p["si"] = {"class_type": "LoadImage", "inputs": {"image": style_img}}
    p["cvl"] = {"class_type": "CLIPVisionLoader", "inputs": {"clip_name": "sigclip_vision_patch14_384.safetensors"}}
    p["cve"] = {"class_type": "CLIPVisionEncode", "inputs": {"clip_vision": ["cvl", 0], "image": ["si", 0], "crop": "center"}}
    p["sml"] = {"class_type": "StyleModelLoader", "inputs": {"style_model_name": "flux1-redux-dev.safetensors"}}
    p["sma"] = {"class_type": "StyleModelApply", "inputs": {"conditioning": ["guid", 0],
        "style_model": ["sml", 0], "clip_vision_output": ["cve", 0], "strength": redux, "strength_type": "attn_bias"}}
    p["ci"] = {"class_type": "LoadImage", "inputs": {"image": comp_img}}
    p["rs"] = {"class_type": "ImageResize+", "inputs": {"image": ["ci", 0], "width": W, "height": H,
        "interpolation": "lanczos", "method": "fill / crop", "condition": "always", "multiple_of": 0}}
    p["depth"] = {"class_type": "DepthAnythingV2Preprocessor", "inputs": {
        "image": ["rs", 0], "ckpt_name": "depth_anything_v2_vitl.pth", "resolution": W}}

def save(p, src, prefix):
    p["dec"] = {"class_type": "VAEDecode", "inputs": {"samples": src, "vae": ["v", 0]}}
    p["save"] = {"class_type": "SaveImage", "inputs": {"images": ["dec", 0], "filename_prefix": prefix}}

def cnet_wf(style_img, comp_img, tag, *, guidance=3.5, scheduler="simple", redux=0.5,
            depth_str=0.7, canny=False, radv=False, hires=False):
    p = {}
    core(p, style_img, comp_img, guidance=guidance, redux=redux)
    pos = ["sma", 0]
    if radv:
        del p["sma"], p["cve"]
        p["radv"] = {"class_type": "ReduxAdvanced", "inputs": {"conditioning": ["guid", 0],
            "style_model": ["sml", 0], "clip_vision": ["cvl", 0], "image": ["si", 0],
            "downsampling_factor": 3, "downsampling_function": "area",
            "mode": "center crop (square)", "weight": 1.0, "autocrop_margin": 0.1}}
        pos = ["radv", 0]
    p["cnl"] = {"class_type": "ControlNetLoader", "inputs": {"control_net_name": "FLUX.1-dev-ControlNet-Union-Pro-2.0.safetensors"}}
    p["cn"] = {"class_type": "ControlNetApplySD3", "inputs": {"positive": pos, "negative": ["neg", 0],
        "control_net": ["cnl", 0], "vae": ["v", 0], "image": ["depth", 0],
        "strength": depth_str, "start_percent": 0.0, "end_percent": 0.8}}
    last = "cn"
    if canny:
        p["cne"] = {"class_type": "Canny", "inputs": {"image": ["rs", 0], "low_threshold": 0.2, "high_threshold": 0.5}}
        p["cn2"] = {"class_type": "ControlNetApplySD3", "inputs": {"positive": ["cn", 0], "negative": ["cn", 1],
            "control_net": ["cnl", 0], "vae": ["v", 0], "image": ["cne", 0],
            "strength": 0.35, "start_percent": 0.0, "end_percent": 0.6}}
        last = "cn2"
    p["msf"] = {"class_type": "ModelSamplingFlux", "inputs": {"model": ["u", 0],
        "max_shift": 1.15, "base_shift": 0.5, "width": W, "height": H}}
    p["lat"] = {"class_type": "EmptySD3LatentImage", "inputs": {"width": W, "height": H, "batch_size": 1}}
    p["ks"] = {"class_type": "KSampler", "inputs": {"model": ["msf", 0], "positive": [last, 0],
        "negative": [last, 1], "latent_image": ["lat", 0], "seed": SEED, "steps": 32, "cfg": 1.0,
        "sampler_name": "euler", "scheduler": scheduler, "denoise": 1.0}}
    final = ["ks", 0]
    if hires:
        p["up"] = {"class_type": "LatentUpscaleBy", "inputs": {"samples": ["ks", 0],
            "upscale_method": "bislerp", "scale_by": 1.5}}
        p["msf2"] = {"class_type": "ModelSamplingFlux", "inputs": {"model": ["u", 0],
            "max_shift": 1.15, "base_shift": 0.5, "width": HW, "height": HH}}
        p["ks2"] = {"class_type": "KSampler", "inputs": {"model": ["msf2", 0], "positive": pos,
            "negative": ["neg", 0], "latent_image": ["up", 0], "seed": SEED, "steps": 32, "cfg": 1.0,
            "sampler_name": "euler", "scheduler": scheduler, "denoise": 0.30}}
        final = ["ks2", 0]
    save(p, final, f"sweeps/{tag}")
    return p

def bfl_wf(style_img, comp_img, tag, guidance):
    p = {}
    core(p, style_img, comp_img, guidance=guidance)
    p["lora"] = {"class_type": "LoraLoaderModelOnly", "inputs": {"model": ["u", 0],
        "lora_name": "flux1-depth-dev-lora.safetensors", "strength_model": 1.0}}
    p["ip2p"] = {"class_type": "InstructPixToPixConditioning", "inputs": {"positive": ["sma", 0],
        "negative": ["neg", 0], "vae": ["v", 0], "pixels": ["depth", 0]}}
    p["msf"] = {"class_type": "ModelSamplingFlux", "inputs": {"model": ["lora", 0],
        "max_shift": 1.15, "base_shift": 0.5, "width": W, "height": H}}
    p["ks"] = {"class_type": "KSampler", "inputs": {"model": ["msf", 0], "positive": ["ip2p", 0],
        "negative": ["ip2p", 1], "latent_image": ["ip2p", 2], "seed": SEED, "steps": 32, "cfg": 1.0,
        "sampler_name": "euler", "scheduler": "simple", "denoise": 1.0}}
    save(p, ["ks", 0], f"sweeps/{tag}")
    return p

def main():
    q = []
    def tagname(r, t): return f"{r.split('.')[0]}x{t.split('.')[0]}"
    for r, t in PAIRS:  # A1: bfl guidance
        for gv in [4.0, 7.0, 10.0]:
            q.append((f"A1-bfl-g{gv:g}-{tagname(r,t)}", bfl_wf(r, t, f"A1-bfl-g{gv:g}-{tagname(r,t)}", gv)))
    for r, t in PAIRS:  # A2: guidance x scheduler
        for gv in [2.5, 3.0, 3.5]:
            for sch in ["simple", "beta"]:
                tag = f"A2-g{gv:g}-{sch}-{tagname(r,t)}"
                q.append((tag, cnet_wf(r, t, tag, guidance=gv, scheduler=sch)))
    for r, t in PAIRS[:2]:  # A3: redux x depth strength
        for rx in [0.4, 0.5, 0.7]:
            for ds in [0.55, 0.7, 0.85]:
                tag = f"A3-rx{rx:g}-ds{ds:g}-{tagname(r,t)}"
                q.append((tag, cnet_wf(r, t, tag, redux=rx, depth_str=ds)))
    for r, t in PAIRS:  # A4: hires
        tag = f"A4-hires-{tagname(r,t)}"
        q.append((tag, cnet_wf(r, t, tag, hires=True)))
    for r, t in PAIRS:  # A5: ReduxAdvanced
        tag = f"A5-radv-{tagname(r,t)}"
        q.append((tag, cnet_wf(r, t, tag, radv=True)))
    for r, t in PAIRS[:2]:  # A6: canny stack
        tag = f"A6-canny-{tagname(r,t)}"
        q.append((tag, cnet_wf(r, t, tag, canny=True)))

    ids = {}
    for tag, prompt in q:
        ids[post(prompt)] = tag
        print("queued", tag, flush=True)
    pending, errors = set(ids), []
    while pending:
        time.sleep(10)
        for pid in list(pending):
            try:
                with urllib.request.urlopen(f"{HOST}/history/{pid}") as r:
                    h = json.loads(r.read())
            except Exception:
                continue
            if pid not in h: continue
            st = h[pid].get("status", {})
            if st.get("completed"):
                pending.discard(pid); print(f"done {ids[pid]} ({len(ids)-len(pending)}/{len(ids)})", flush=True)
            elif st.get("status_str") == "error":
                pending.discard(pid); errors.append(ids[pid])
                msgs = [m for m in st.get("messages", []) if m[0] == "execution_error"]
                print(f"ERROR {ids[pid]}: {(msgs[-1][1].get('exception_message','?') if msgs else '?')[:300]}", flush=True)
    os.makedirs(OUT_ROOT, exist_ok=True)
    src = os.path.join(COMFY_OUT, "sweeps")
    for f in sorted(os.listdir(src)):
        shutil.copy(os.path.join(src, f), os.path.join(OUT_ROOT, f.split("_")[0] + ".png"))
    print(f"COLLECTED {len(os.listdir(OUT_ROOT))} sweep outputs", flush=True)
    if errors:
        print("ERRORS:", errors); sys.exit(1)
    print("SWEEPS COMPLETE", flush=True)

if __name__ == "__main__":
    main()