## Instructions - Write prompts as Subject + Action + Environment + Light + Lens + Style: one sentence each, no purple prose. - Lock consistency with reference image + seed: same seed + same ref for character, vary only pose/prompt delta. - Control with negative prompts and weights: (blurry:1.2), bad hands, extra fingers, watermark — test one at a time. - Pick local vs cloud deliberately: ComfyUI local for privacy/iterations, Midjourney/Runway cloud for speed/quality. - Storyboard first: 6-panel script with shots, duration, dialogue — generate stills before any video render. - Upscale last: base 512/768, Hires-Fix 2x, then Topaz/Real-ESRGAN to 4K — never prompt at 4K directly. - Disclose AI use: label client work AI-assisted, check model license for commercial use before selling. - Respect copyright: no artist-name mimic for sale, no real-person likeness without consent, keep prompt logs. - Iterate at low steps (20-25) for composition, then high steps (40-60) + refiner for final only to save GPU. - Direct video with image-to-video: strong keyframe still + 2s motion brush + camera move — not text-to-video chaos. - Save LoRAs for recurring characters/products: 20-30 captioned images, rank 8-16, test at 0.7 weight first. - Archive runs: prompt/seed/model/version/negative saved with PNG info or ComfyUI JSON for reproducibility. ## Examples - Example 1: Product shot: matte bottle on stone, softbox left, 85mm f2.8, studio — seed 42, upscale 4K for Amazon. - Example 2: Consistent mascot: ref sheet front/side, seed locked 1234, 10 poses via prompt delta only. - Example 3: Talking promo: Midjourney still + Runway image-to-video lip move 4s + CapCut captions + disclosure. - Example 4: Local anime: ComfyUI SDXL + LoRA 0.8, Hires-Fix, 30-min batch 100 variants, pick 3 finals. ## Workflows - Workflow A — Image final: 1. Prompt draft 2. Low-step batch 3. Pick + seed lock 4. Refine 5. Hires 6. Upscale 7. Log. - Workflow B — AI video: 1. Script 2. Keyframe stills 3. Image-to-video 4s 5. Upscale 6. Edit + sound 7. Disclosure. - Workflow C — Character: 1. Ref sheet 2. Train LoRA 3. Test poses 4. Lock seed 5. Batch scenes 6. Style guide. - Workflow D — Client delivery: 1. License check 2. Prompt log 3. Human retouch 4. Label AI 5. 4K master + social cuts. ## Templates - Image-gen prompt template: ``` Subject: [28yo hiker, red jacket, freckles, reference IMG_01] Action: [standing on cliff, arms open, jacket flutter] Environment: [Dolomites sunset, clouds below] Light/Camera: [golden hour rim, 85mm f1.8, shallow DOF] Style: [photoreal, Kodak Portra 400 --ar 16:9 --style raw] Negative: [blurry, cartoon, extra fingers, watermark, oversaturated] Seed: 4242, Steps: 30, CFG: 6.5, Model: SDXL v1.0 + Detail LoRA 0.7 ``` - Video shot list: Shot01 0-3s wide still + slow push-in | Shot02 3-6s close-up + pan right | Audio VO + music -14 LUFS. - ComfyUI checklist: Checkpoint + VAE + Hires-Fix 1.5x + KSampler dpm++2M + Save JSON with seed. - Disclosure line: Created with AI assistance (Midjourney/ComfyUI) + human edit — commercial license verified {date}. ## Tools - Midjourney / DALL-E 3 / Flux — fastest quality stills; --cref --cw for character reference locking. - Runway / Pika / Luma — image-to-video with camera control and motion brush for ads. - ComfyUI + SDXL/Flux local — node graph, LoRA training, private NSFW-safe iteration on RTX 3060+. - Topaz Photo/Video AI + Real-ESRGAN — 2x-4x upscale that beats in-model Hires for print. - Civitai + HuggingFace — vetted checkpoints/LoRAs with license tags before commercial use. - CapCut / Resolve — assemble AI clips, auto-captions, loudness normalize for TikTok/YouTube. ## Advanced - Train product LoRA: 25 phone photos varied angles, BLIP caption, 2000 steps rank 16, trigger token SKU_X. - Chain ControlNet: depth + openpose + canny to hold composition while changing style/outfit. - Batch with XYZ plot: seeds 1-20 x CFG 5-8 to find sweet spot in one overnight ComfyUI queue. - Fix hands/faces with inpaint + ADetailer at 0.4 denoise — do not regenerate whole image. - Estimate cost: cloud $0.05/image vs local $0.01 power — switch to local over 500 images/month. - Watermark provenance with C2PA/Content Credentials + keep raw prompt JSON for disputes. - Keep a swipe file of 50 winning prompts with seeds; remix lighting/subject instead of starting blank. - Test every LoRA at 0.5/0.7/1.0 weights on same seed before committing to batch render. ## Documentation - Midjourney docs — --cref --sref --seed consistency and style raw tuning guide. - ComfyUI Wiki — manager, custom nodes, Hires-Fix and ControlNet wiring examples. - Runway/Pika prompting — camera moves, motion brush strength, lip-sync limits. - Civitai LoRA training — dataset curation, captioning, overfitting checks. - Copyright Office AI guidance — disclosure, human-authorship threshold, commercial risk notes.