#!/usr/bin/env python3 """Build Boogu LoRA (boonude) workflow JSONs in ComfyUI UI format. Graph (Base variant): UNETLoader -> LoraLoader(model,clip) -> ModelSamplingAuraFlow -> KSampler CLIPLoader -> LoraLoader.clip -> CLIPTextEncode(+/-) -> KSampler VAELoader -> VAEDecode ; EmptyLatentImage -> KSampler ; SaveImage """ import json BASE_UNET = "boogu_image_base_fp8_scaled.safetensors" TURBO_UNET = "boogu_image_turbo_fp8_scaled.safetensors" LORA = "boonude.safetensors" CLIP = "qwen3vl_8b_fp8_scaled.safetensors" VAE = "ae.safetensors" PROMPT = "A beautiful portrait of a woman, soft studio lighting, detailed skin texture, professional photography" # Official Base negative prompt from the Boogu project's own workflow (boogu_base_t2i.json) OFFICIAL_NEGATIVE = ("(((deformed))), blurry, over saturation, bad anatomy, disfigured, poorly drawn face, " "mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), fused fingers, " "messy drawing, broken legs censor, censored, censor_bar") def make_workflow(unet, prefix, sampler_widgets, auraflow, negative=OFFICIAL_NEGATIVE): """Return (workflow_dict). Nodes are defined with explicit edges; links are auto-numbered.""" # node specs: id -> (type, widgets, inputs[(name,type)], outputs[(name,type)]) specs = [] edges = [] # (from_node, from_slot, to_node, to_slot, type) def add(ntype, widgets, inputs, outputs, pos, size=(300, 100)): nid = len(specs) + 1 specs.append((nid, ntype, widgets, inputs, outputs, pos, size)) return nid unet = add("UNETLoader", [unet, "default"], [("unet_name", "COMBO")], [("MODEL", "MODEL")], (0, 0)) clip = add("CLIPLoader", [CLIP, "boogu", "default"], [("clip_name", "COMBO")], [("CLIP", "CLIP")], (0, 300)) lora = add("LoraLoader", [LORA, 1.0, 1.0], [("model", "MODEL"), ("clip", "CLIP")], [("MODEL", "MODEL"), ("CLIP", "CLIP")], (360, 0)) edges.append((unet, 0, lora, 0, "MODEL")) edges.append((clip, 0, lora, 1, "CLIP")) cur_model = lora cur_clip = lora if auraflow: af = add("ModelSamplingAuraFlow", [3.16], [("model", "MODEL")], [("MODEL", "MODEL")], (720, 0)) edges.append((lora, 0, af, 0, "MODEL")) cur_model = af vae = add("VAELoader", [VAE], [("vae_name", "COMBO")], [("VAE", "VAE")], (0, 480)) pos = add("CLIPTextEncode", [PROMPT], [("clip", "CLIP")], [("CONDITIONING", "CONDITIONING")], (400, 300)) edges.append((cur_clip, 1, pos, 0, "CLIP")) neg = add("CLIPTextEncode", [negative], [("clip", "CLIP")], [("CONDITIONING", "CONDITIONING")], (400, 520)) edges.append((cur_clip, 1, neg, 0, "CLIP")) latent = add("EmptyLatentImage", [1024, 1024, 1], [], [("LATENT", "LATENT")], (760, 480)) ksampler = add("KSampler", sampler_widgets, [("model", "MODEL"), ("positive", "CONDITIONING"), ("negative", "CONDITIONING"), ("latent_image", "LATENT")], [("LATENT", "LATENT")], (1100, 0)) edges.append((cur_model, 0, ksampler, 0, "MODEL")) edges.append((pos, 0, ksampler, 1, "CONDITIONING")) edges.append((neg, 0, ksampler, 2, "CONDITIONING")) edges.append((latent, 0, ksampler, 3, "LATENT")) decode = add("VAEDecode", [], [("samples", "LATENT"), ("vae", "VAE")], [("IMAGE", "IMAGE")], (1460, 100)) edges.append((ksampler, 0, decode, 0, "LATENT")) edges.append((vae, 0, decode, 1, "VAE")) save = add("SaveImage", [prefix], [("images", "IMAGE")], [], (1820, 100)) edges.append((decode, 0, save, 0, "IMAGE")) # assign link ids links = [] for i, (fn, fs, tn, ts, t) in enumerate(edges, start=1): links.append([i, fn, fs, tn, ts, t]) # build nodes with consistent per-slot link ids in_link = {} # (node, slot) -> link id out_link = {} # (node, slot) -> list of link ids for i, (lid, fn, fs, tn, ts, t) in enumerate(links, start=1): in_link[(tn, ts)] = lid out_link.setdefault((fn, fs), []).append(lid) nodes = [] for nid, ntype, widgets, inputs, outputs, pos, size in specs: node = { "id": nid, "type": ntype, "pos": list(pos), "size": list(size), "flags": {}, "order": nid - 1, "mode": 0, "inputs": [ { "name": name, "type": t, "link": in_link.get((nid, i)), **({"widget": {"name": name}} if t == "COMBO" else {}), } for i, (name, t) in enumerate(inputs) ], "outputs": [ {"name": name, "type": t, "links": out_link.get((nid, i), []), "shape": 3} for i, (name, t) in enumerate(outputs) ], "properties": {"Node name for S&R": ntype}, "widgets_values": widgets, } nodes.append(node) return { "last_node_id": len(specs), "last_link_id": len(links), "nodes": nodes, "links": links, "groups": [], "config": {}, "extra": {"ds": {"scale": 0.5, "offset": [0, 0]}}, "version": 0.4, } base_wf = make_workflow( BASE_UNET, "Boogu_lora_base", # official Base recipe: 50 steps / cfg 4.0 (from boogu_base_t2i.json) [123456789012345, "randomize", 50, 4.0, "dpmpp_2m", "simple", 1.0], auraflow=True, ) turbo_wf = make_workflow( TURBO_UNET, "Boogu_lora_turbo", [987654321098765, "randomize", 4, 1.0, "lcm", "sgm_uniform", 1.0], auraflow=False, ) with open("boogu_lora_t2i_base.json", "w", encoding="utf-8") as f: json.dump(base_wf, f, ensure_ascii=False, indent=1) with open("boogu_lora_t2i_turbo.json", "w", encoding="utf-8") as f: json.dump(turbo_wf, f, ensure_ascii=False, indent=1) # sanity checks for name, wf in (("base", base_wf), ("turbo", turbo_wf)): ids = {n["id"] for n in wf["nodes"]} assert all(l[1] in ids and l[3] in ids for l in wf["links"]), name # every node input link must reference an existing link id all_lids = {l[0] for l in wf["links"]} for n in wf["nodes"]: for inp in n["inputs"]: if inp["link"] is not None: assert inp["link"] in all_lids, f"{name} node{n['id']} input {inp['name']}" print(f"{name}: nodes={len(wf['nodes'])} links={len(wf['links'])} OK")