Image generation

Image models, control, editing, inpainting, identities, consistency and upscaling

• articleHow do I pin where text goes in a generated graphic using structured controls (FLUX 3 layout boxes, Recraft V3 text_layout), and which scripts and characters do they support?
  Problem. Prompt-only placement drifts; some APIs accept coordinates. Conditions. FLUX 3 Image (BFL) and Recraft V3 / V3 Vector. Method (documented). FLUX 3: a layout prompt is a caption plus a JSON array of elements, each with an id, a bbox and a desc; give each text block its own row, quote the words, and state case, colour and typeface character in the desc; the caption cites the element by its id. A \n inside the quoted string is a line break; 'arranged vertically' sets reading direction. BFL's examples include Japanese and English text on one poster. Recraft: text_layout is an array of {text, bbox} where each item is one word and bbox is four relative points (0,0 top-left, 1,1 bottom-right); coordinates beyond 0 to 1 crop the shape; the field works on V3 and V3 vector models only (the V4/V4.1/V2 columns in the parameter table show no support); the text characters allowed are uppercase Latin letters, digits, basic punctuation, a handful of Latin-extended, Greek and Cyrillic capit...
• articleWhat does Recraft document for SVG output (vector models, Vectorize, SVG background removal), and what is it silent about regarding editability and tracing?
  Problem. 'Vector' from an AI tool may be genuine shapes or a trace of a raster; designers need to know which before handing off. Conditions. Recraft API docs, 23 Oct 2026. Method (documented). Vector output is always SVG; vector models include recraftv4_1_vector, recraftv4_vector, recraftv3_vector, recraftv2_vector and Pro variants. Vectorize (POST /v1/images/vectorize) converts PNG, JPG or WEBP up to 20 MB and 16 MP into an SVG the docs describe as clean, editable and made of shapes, with shape_stacking set to hierarchical (default) or cut_out. removeBackground accepts SVG in and returns SVG out. Text can be placed by text_layout on V3 vector models (C3-B593-02). What the docs do not say [inference from absence in the pages read]: whether the _vector generation models draw native shapes or generate raster then trace; node counts, path complexity, gradients or text as live text versus outlines; font handling. Acceptance check [inference]: open the SVG in a vector editor; count paths...
• articleWhat style, preset and seed controls do Ideogram, Recraft and FLUX 3 document for repeatable image sets, and what is missing (e.g. FLUX 3 has no seed)?
  Problem. Repeating a look across a series: which knob does each tool give? Conditions. Docs of 23 Oct 2026. Method (documented). Ideogram 3.0 generate: seed (reproducible generation when set), style_reference_images (max 10, max 25 MB each; JPEG, PNG, WEBP), style_codes (8-character hex), style_type and style_preset (the preset cannot be combined with codes or references), color_palette, and magic_prompt (automatic prompt rewriting, default auto). Ideogram 4.5 edit: images to edit (max 5), the first is the one being edited and any others are extra references. Recraft: random_seed for reproducibility plus style references (C3-B593-07). FLUX 3 Image: the prompt carries everything; there is no negative_prompt, guidance or seed field, and bounding boxes go at the end of the prompt. Implications [inference]. (1) Seed reuse gives repeatability only for an unchanged prompt and settings on one model version; it does not carry a look to a different prompt, so for series work use style refere...
• articleWhich palette and exact-colour controls do FLUX, Recraft and Ideogram document, and how should I verify that a generated image kept the brand colours?
  Problem. Brand work needs specific colours; prompts alone drift. Conditions. FLUX.2 / FLUX 3 (BFL), Recraft API, Ideogram 3.0 API, docs of 23 Oct 2026. Method (documented). BFL: hex codes name exact colours in FLUX.2 prompts (docs claim brand-guideline matching 'with precision'); the FLUX 3 edit docs say to tie each hex code to one named object (for example only one bird in the middle) so the model knows where the colour goes. Recraft: controls.colors is an array of up to 10 {rgb:[r,g,b]} objects with an optional weight 0 to 1 (set for all or none, sum at most 1) and controls.background_color; the docs call them 'preferable colors'. A palette can also be stored with a style (Create style). Ideogram 3.0: a color_palette parameter exists (schema not expanded in the page read), plus style_codes (8-character hex codes) and style_preset; the preset cannot be combined with codes or references. Verify [inference]. Sample the output in sRGB, cluster to the dominant five to eight colours, an...
• articleWhen should I train a style LoRA instead of using reference images, and what do BFL's FLUX.2 klein guide say about dataset size, captions and trigger words?
  Problem. A brand look must hold across hundreds of images, offline or at low cost. Conditions. FLUX.2 [klein] LoRA training with an ai-toolkit-style config in BFL's guide; check the klein model licence before commercial use (not read here). Method (documented). The worked example trains a style LoRA on a fully synthetic dataset (images generated with FLUX.1 [dev] and illustrative LoRAs): 27 training images with caption files. Captions should describe only what is in the image and no style descriptors, so the model learns style from the pixels; a unique trigger word in every caption (example GFX_IMPR5N) activates the style at inference without polluting the base vocabulary. Training length guidance in the guide: style LoRAs (10 to 20 images) 1200 to 1800 steps, character LoRAs (10 to 15 images) 800 to 1200, medium datasets (20 to 40 images) 1200 to 2000, large datasets (50+ images) 2000 to 3000; the example config uses 3000 steps and inference with 8 steps for a soft painterly look. ...
• articleHow do Recraft's style references and style_id work for keeping a set of images on-brand (1-10 references, flexible vs precise match), and what are the rules?
  Problem. Producing many images that share one visual identity. Conditions. Recraft API V2/V3/V4/V4.1 (not V4.1 Flash), docs 23 Oct 2026. Method (documented). A custom style is built from one to ten reference images (PNG, JPG, WEBP; 64 MB total, each file under 20 MB) with no training; references can come from other generators, branding materials or existing assets. Use it by style_id (create once with Create style) or by attaching style_references / style_reference_urls to a generation, which creates a private style and returns its style_id for reuse; sending both is rejected. V4 and V4.1 styles have a match value: flexible (default) matches the general vibe, precise follows the style meticulously, including rendering technique, colour, composition and lighting; style_match in a request overrides the stored value. The dedicated V4 Styles models always require a style. Guidance: a single reference is enough to start; similar images sharpen the match and diverse ones widen the range. ...
• articleWhat do OpenAI and Google docs say about keeping characters and brand elements consistent across generations, and how many reference images can I give?
  Problem. Logos, mascots and products must repeat accurately across images. Conditions. OpenAI GPT Image docs and Gemini (Nano Banana) docs, 23 Oct 2026. Method (documented). OpenAI lists consistency as a limitation: the model may occasionally struggle to keep recurring characters or brand elements consistent across multiple generations; it also lists composition control (difficulty placing elements precisely in structured or layout-sensitive compositions). Google's model summary describes the premium Gemini 3 Pro Image model as offering accurate brand consistency and its Nano Banana 2.1 as improving multi-turn consistency (vendor claims); the Gemini reference table gives up to 4 or 5 character images for character consistency depending on the model. Use [inference]: treat logos as assets, not generations: place the real logo file (vector or high-res PNG) in the composite after generation; use the model for backgrounds and scenes; if a mascot must recur, use the vendor's reference-im...
• articleWhat generation metadata is lost when a generated image is re-encoded, converted or uploaded, and how should an agent preserve it?
  Partially verified: the curator re-read none of the cited sources; S8, S9, S2, S1, S4, S3 rest on the producer's pinned reading. Embedded settings live in ancillary PNG text chunks (or EXIF/tags) that decoders ignore, so they exist only while a PNG-aware tool copies them. The PNG specification says ordinary image editors usually discard unrecognized information; the tools' own code writes nothing when their disable switches are set and writes different containers per format. Practical rule (synthesis): extract and store the metadata together with a hash of the original bytes before any resize, convert, screenshot or upload, because the sources inspected give no way to recover it afterwards. Behaviour of specific hosting sites was not verified.
• articleAccording to ZEISS's lens-division paper, what actually controls depth of field (f-number, imaging scale, format), and how should that change how I prompt for 'shallow depth of field' and lens length?
  Problem. Prompts mix 'shallow depth of field', 'f/1.4' and '85mm' as if they were interchangeable. Conditions. Any photoreal image/video prompt that names lens or aperture. Source statements (H. H. Nasse, Carl Zeiss Camera Lens Division, 2010). Depth of field depends on the acceptable blur and so rests on arbitrary specifications; it does not stop sharply but changes continuously. In object space it is (almost) independent of focal length when compared at the same imaging scale (subject size on the sensor), meaning different focal lengths shot from correspondingly different distances give similar depth of field; this stops holding at very large distances. Differently, perspective and the maximum blur of the distant background differ: background blur is lower for shorter focal lengths than for longer ones. Describing a lens by its 35 mm-equivalent focal length is inconsistent unless the aperture is converted too; equivalent f-numbers scale with format size, so small formats have litt...
• articleThe invented-brand product spot: packshot with rendered label text, then a 15 s multi-cut commercial (Apr-Jun 2026 creator recipe)
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Recurring recipe in Artlist (2026-04-14) and Creative Fabrica (2026-06-02) videos: (1) generate a packshot of a product with an INVENTED brand name rendered on the label (Nano Banana Pro or GPT Image 2; e.g. a soda can or chocolate bar named after the creator), 16:9, one image; (2) image-to-video with Seedance 2.0: "Fast multiple cuts and cinematic camera movement. Premium beverage commercial. No music, only sound effects", 720p, 15 s, audio on; (3) iterate background by prompt ("place the product on a beach"). He reports good product and label consistency. Version-claim drift to note: in April he said Seedance 2.0 combines "up to 12 references"; later videos cite 50 re...
• articleWorldbuilding in service of story, and using concept-art pushes to feed writing (Sanderson, Dos Santos, Stewart panel)
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. DSNX 2025 panel (published 2026-01-23). Practical points: (1) Ratio and order - Sanderson says he now writes about 10,000 words of pre-material per 100,000 finished words and works better alternating between story and world document than building a huge bible first; his six-month pure-worldbuilding binder for Way of Kings Prime was mostly unused and the published book restarted from character arcs. (2) Test an idea for story, not coolness: a striking setting (seas of spores/pollen) was shelved for years because it had no inherent danger or conflict; it worked once tied to a dangerous magic mechanism. (3) Concept-art pushes - a group brainstorm over a fixed period, often...
• articleA four-stage control ladder for AI video: one-line prompt, LLM-expanded multi-shot prompt, edited storyboard grid, then saved character/location assets
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Youri van Hofwegen (Higgsfield link, 2026-08-11) holds model (Seedance 2.0), 1080p, 16:9 and the idea constant and changes only workflow: (1) one plain sentence - fast for testing, but camera, sound and look are the model's choices; (2) an LLM (Claude) writes a detailed three-shot prompt - deliberate cuts and camera, though a requested dialogue line was dropped; (3) a three-panel storyboard image from GPT Image 2, fixing a wrong detail (paved vs dirt road) by editing only the affected panels, then giving the storyboard plus an LLM prompt written from it - output followed the board shot-for-shot; (4) saved assets: a character sheet made from a COLLAGE of real photos (mul...
• articleHow do I outpaint a still with FLUX Fill or SDXL inpainting in diffusers (strength, mask blur, resolution) and which licences apply?
  Problem. Extending a frame beyond its edges while keeping the original pixels. Conditions. diffusers FluxFillPipeline or an SDXL inpainting checkpoint. Method. (1) FLUX.1 Fill [dev]: the docs say the pipeline supports inpainting and outpainting and does not take a strength input; pass prompt, image, mask_image, height, width and max_sequence_length=512; the BFL doc says the mask must be a black-and-white image the same size as the conditioning image. For outpainting (inference): paste the source on a larger canvas, make the mask white over the new area, and set height/width to the canvas size. (2) SDXL Inpainting 0.1: guidance_scale 8.0, 15-30 steps, and strength=0.99 - the card says to keep strength below 1.0; checkpoint is openrail++. (3) General inpaint tuning from the diffusers guide: strength is how much noise is added to the base image, higher gives more difference from the original; blur_factor on the mask processor softens the seam (example 33); padding_mask_crop crops aroun...
• articleWhat does ComfyUI-AdvancedLivePortrait offer, what are its dependency pins, and what is unclear about its licence?
  Problem. Editing facial expressions or driving a portrait from video inside ComfyUI. Conditions. PowerHouseMan/ComfyUI-AdvancedLivePortrait commit 3bba7329 (last push 2024-08-21; not archived); GitHub reports no licence; pyproject version 1.0.0 declares license = {file = "LICENSE"}, but no LICENSE file was found at that commit in my fetch. Method. README features: edit expressions in photos, insert expressions into videos, make animations from multiple expressions, extract expressions from sample photos, real-time preview, and (update 2024-08-21) video creation without a source video and face tracking in the source video. A 'command' syntax links motions as [Motion index] = [Changing frame length] : [Length of frames waiting for next motion], motion index 0 being the source image; expressions can be saved/loaded to output/exp_data. It uses a model converted by kijai (from ComfyUI-LivePortraitKJ) and the original LivePortrait. Requirements include numpy>=1.26.4, opencv-python-headles...
• articleWhat do the InsightFace repositories say about the licence of buffalo_l, antelopev2 and other pretrained models versus the MIT-licensed library, and who must be contacted for commercial use?
  Problem. IPAdapter FaceID, face swap and portrait nodes in ComfyUI depend on insightface models. Conditions. deepinsight/insightface commit c6c3af7b (2026-10-19). Cross-checks an earlier round's licence item without repeating it. Method. The main README states the code is MIT with no limit on academic or commercial use, but training data with annotations 'and the models trained with these data' are for non-commercial research only, and this applies both to manual downloads and to auto-downloads through the Python library. It lists contacts for licensing: inswapper-series face swap models (inswapper_128.onnx etc.), the open-sourced face recognition packages such as buffalo_l (a separate recognition-oss-pack contact), and the InspireFace SDK. The model_zoo README opens with 'ALL models are available for non-commercial research purposes only'; the python-package README repeats that the pretrained models are non-commercial research only, whether downloaded automatically or manually, and...
• articleWhat does a "copy a viral Short's structure" AI Shorts workflow look like, and what policy risks does it carry?
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Youri van Hofwegen (2026-07-09, OpenArt affiliate): find a high-view Short in any niche, pull its transcript (third-party transcript site), ask Claude to rewrite a new script for your niche "in the same style, length, and tone", generate an ElevenLabs v3 voice-over (filter voices by accent/gender/"conversational"), have Claude write image prompts tailored to the target image model (GPT Image 2, 1:1, high quality) for meme stills, lay them over muted "copyright-free" gameplay footage in CapCut with auto-captions. He claims ">90%" of beginners get under 1,000 views because they chase originality, unsourced. Risks under YouTube rules: templated, mass-produced or near-ident...
• articleHow can an instruction image-edit model (Qwen Image Edit + next-scene LoRA) build a consistent storyboard/scene shot by shot, and how do you keep backgrounds consistent?
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Mickmumpitz (2025-11-01): Qwen Image Edit (GGUF Q5 fits 16 GB; 4-step Lightning LoRA with CFG 1) composes characters into a location from references; lovis93's next-scene LoRA (MIT on HF) generates the following shot from the previous image. Lessons: prompting camera moves ("camera behind the man, over-the-shoulder") failed (duplicate man, no rotation) until he described where things are in space ("half the photo obstructed by the man's back, window to the right, bar behind the woman to the left"); re-attach the original character image when likeness drifts; optional pose references; to keep backgrounds consistent he trained a LoRA on 20 real 360-degree images so the mo...
• articleWhat does ComfyUI-GGUF let me do with quantised diffusion models (Flux, SD3.5), what limits does it state, and does a GGUF conversion change the licence?
  Problem. A diffusion transformer does not fit in VRAM, or a team wants a smaller file to move between machines, and must know what the GGUF route changes. Conditions. city96/ComfyUI-GGUF at commit 6ea2651e (last commit 2026-01-12; code licence Apache-2.0 text). Targets transformer/DiT models (the README names Flux and SD3.5 Large quants) and a quantised T5 text encoder. Method. (1) Put .gguf files in ComfyUI/models/unet and replace the stock 'Load Diffusion Model' with the 'Unet Loader (GGUF)' node (category bootleg). (2) For the T5 encoder use the *CLIPLoader (gguf) nodes, which accept gguf and regular safetensors/bin. (3) Install the gguf Python package. (4) Keep ComfyUI recent enough for custom ops. (5) Treat LoRA use as experimental: the README says the built-in LoRA loaders 'should work'. What it changes. The README claims DiT models seem less affected by quantisation than conv2d UNets, which is why low-bit quants are viable on low-end GPUs; it gives no quality measurements, an...
• articleVendor craft guidance for multi-shot AI films: in-prompt cuts, world bible, audio motif, one beat per shot (BFL cookbook)
  Black Forest Labs cookbook (with notebook): a generation is capped at 20 s, so build length two ways. (1) Cuts inside one generation: label SHOT ONE / HARD CUT / SHOT TWO; consecutive shots must differ strongly - scale jumps (macro -> wide -> aerial) read as deliberate cuts while similar angles blur into one drifting take; bind the shots with one audio idea (a continuous bed, with a sound source scaled per shot rather than restarted); keep to 2-3 shots per generation. (2) A film as a shot list generated concurrently (rate limits are per concurrent generation) and cut with ffmpeg; a miss costs one shot, and you keep editorial timing. Craft rules: a one-paragraph WORLD BIBLE (place, palette, light, grade) pasted word-for-word into every shot prompt - consistency comes from consistent text, so do not paraphrase; a named musical motif staged per shot (enters, warmer, builds, resolves); one beat per shot. Result URLs expire about 2 h after Ready, so download promptly. Treat moderation st...
• articleKrea 2 open weights (Raw and Turbo, Jul 2026): free commercial use only below $1M company revenue, with naming, content-filter and AI-disclosure duties
  Closes the Krea licence follow-up. Krea 2 is released as "Krea 2 Raw" and "Krea 2 Turbo" weights (HF cards updated 2026-07-24) plus style LoRAs, under the Krea 2 Community License v.1 (22 June 2026); Krea's realtime video model is separately Apache-2.0 (R2). Key terms: Commercial Use of the model, derivatives OR OUTPUTS is permitted only if you and all affiliates have under $1,000,000 total annual revenue (trailing 12 months, all sources); at or above that you need an Enterprise License before commercial use, and must stop immediately if you cross the threshold mid-use. Distributing the model or a product containing it requires passing on the licence, prefixing any derived model name with "Krea" (e.g. "Krea 2 [Name]"), and shipping a Notice file with the given attribution (and stating modifications). Deployers must implement reasonable content filters (examples given: open classifiers, moderation APIs, human review) and must disclose AI generation where law or platform policy requir...
• articleMidjourney in Aug-Oct 2026: V8.2 edit model (instruction edits, 4 image references, in/outpainting), "Thinking Mode" reruns, shared folders
  Primary update feed. 27 Aug 2026: first V8.2 image EDIT model opened to everyone - edit images by instruction, generate from up to 4 image references at once ("replacing omni-reference"), inpainting and outpainting, works with personalization, moodboards and style refs (which may need extra prompt direction); usable via prompt-bar attachments, lightbox Edit, the Edit tab, or --edit on Discord; image quality update on 28-29 Aug. Sep 2026 (alpha site): style live-previews using fast models, default parameters per user/folder, improved in/outpainting that changes only selected pixels so repeated edits do not degrade the image, Korean UI. 7-8 Oct 2026: "Thinking Mode" test - rerun a V8.2 job "with thinking": Midjourney inspects the image for misses against the prompt (wrong objects, layout, anatomy, text) and regenerates; reported to help prompt accuracy, typography and coherence; shared folders with collaborators. Not found in the feed: a new Midjourney video model release in this wind...
• articleSeedance 2.5 reference discipline and "acting with your phone": describe and group references, add a consistency/negatives block, imperfection cues for realism, and drive performances from iPhone footage
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Dan Kieft (Aug-Sep 2026, condensed reads). Reference discipline (R1): with many reference slots, describe each reference in the prompt and GROUP references that show the same person/object from different angles (separate angle images rather than one multi-angle sheet); state the host's subject rule (he cites "one to eight distinct subjects" across subject references); add a "keep consistent" block (format clarity e.g. handheld phone, colour and grain, lighting, shadow, mood) and explicit exclusions ("no brand logos, extra people, text on screen"). Realism (R2): make the still cinematic first (e.g. 35 mm, f/1.8, shallow depth of field, fine film grain, muted desaturated ...
• articleCredit-saving direction habits (Sep 2026): motion-only video prompts, don't name things you don't want rendered, and let an LLM "director" write second-by-second blocking
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Youri van Hofwegen. Higgsfield production (R1, sponsored): wasted credits are "almost never the model's fault"; generate the character sheet and environment once (GPT Image 2 for photorealism), then drop BOTH in as references for each shot; when a shot already has a start frame, the video prompt "only has to cover motion and camera"; use lens/angle language deliberately (e.g. floor-level, steep up-angle, 24 mm wide to make a subject look powerful); and avoid naming objects you do not want - "anything you name as an object is something the model will try to render" (naming "the camera" can put a camera in frame). Settings he used: Seedance 2.5 at 10 s, 1080p, 16:9; also ...
• articlePrompt-based video modification (PixVerse Modify) and multi-reference self-insertion (WeryAI + Nano Banana Pro + Seedance 2.0)
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. PixVerse Modify (sponsored, 2026-04-08): on an existing clip choose Swap (mask an object, upload a replacement image - e.g. a perfume bottle replaced by a soda bottle), Add ("add an airplane flying over the scene"), Remove (mask + "remove label") or Restyle (3D animation, 2D animation, comic, ink wash or custom) while motion, lighting, perspective and camera move are kept - he presents these as natural-language edits without special syntax. WeryAI (sponsored, 2026-04-16): generate four candidate frames with Nano Banana Pro, then image-to-image with two references (the chosen frame + a photo of himself) and the constraint "Preserve face and hair 100%", then Seedance 2.0 ...
• articleAn end-to-end AI short-film method: reference sets per character, object and location continuity, three prompt formats, 480p drafts, voice consistency and motion transfer
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Dan Kieft (OpenArt-sponsored course, 2026-06-25; says he spent over $1,000 in credits on the short). Pre-production: per prominent character use three references - a character sheet, a face close-up, and a skin/hair detail sheet - generated with GPT Image 2 at high quality (preferred over Nano Banana Pro for likeness); use a GREY not white sheet background because white transferred as overexposure in his A/B; real photos of a real subject beat AI sheets. Lock locations with a mood board/master location and treat props as references too: a TV that vanished mid-shot was fixed only by adding it as a reference. He uses an LLM (Claude) with his own Markdown "prompt guide" fi...
• articleStyle frames first, prompt like a cinematographer; and the self-initiated brand concept film as a portfolio format
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Dreamina video (promotional, 2026-07-14): lock a style frame before animating because it drives all later animations; generate it with GPT Image 2 from a reference photo of the creator; write prompts "like a cinematographer" - camera angle, lighting, atmosphere, lens, colour grade, mood, environment, subject detail; use multiple reference images and prompt tweaks to converge on a style; then animate in Seedance 2.0 with omni-reference (image, video, audio). Lamborghini concept film (2026-08-14): a self-initiated, no-brief commercial for a real car brand mixing traditional VFX and AI; the transcript is mostly the voiceover and music (no technical breakdown audible). Righ...
• articleExtending past a model's clip limit by re-seeding from exported highlight frames (creator trick, Seedance 2.0)
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Mikkel Lassalle (Loova referral, 2026-04-28): to go beyond the 15 s Seedance 2.0 limit, download the clip, export a few highlight frames in any editor, and feed them back as multiple image references with a short action prompt (multiple cuts, camera movement, sound effects, grade) to get continuation clips that are stitched in the edit. Also: create named reusable "characters" from an uploaded image in the reference-to-video tool, then remove raw image tags once characters are tagged. Packaging note: the title promises a comparison with traditional VFX, but the transcript contains no such comparison (title/content mismatch; relevant to title-honesty work). Partly supers...
• articleA creator VFX workflow for Seedance 2.5: character sheets, meta-prompting and tagged references for 30 s shots
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Mikkel Lassalle (OpenArt influencer link, 2026-10-07). Workflow: (1) build assets first - character sheets on a plain grey background with one close-up and one full-body shot (he recommends GPT Image 2 at medium quality, 16:9, 2K; for his own likeness the prompt says preserve identity 100% and changes clothes); also object sheets (a vintage plane) iterated by prompt edits; (2) "meta prompting": paste a template prompt into ChatGPT with a short story idea and let it expand a detailed 30 s Seedance prompt; (3) in Seedance 2.5 select the sheets as references, set duration up to 30 s, and tag each reference to the matching role in the prompt (main character, companions, air...
• articleHow do creators make AI video look like real camera footage (prompt formula, imperfection, camera behaviour) and pick models for realism?
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Dan Kieft (2026-10-16, Higgsfield-sponsored): for stills, test the same prompt across several image models because each is trained differently (in his comparisons GPT Image 2.5 "Flare", FLUX 2 Pro, Nano Banana Pro/2.1, Seedream 5.0 and Higgsfield Soul each won some cases; FLUX 3 looked plastic for stills; text rendering failed on Soul 2 and FLUX 2 Pro). His "CTLT" prompt order: Camera (film stock/lens/framing, e.g. 35 mm film, 12 mm fisheye, tilted horizon, slight motion blur), Tone (mood/emotional energy, e.g. "we weren't meant to see this"), subject/action, Lighting (e.g. harsh on-camera flash), Texture (physical surface details); style labels such as disposable-camer...
• articleHow should you structure prompts and references for Kling 4.0 to get 30-second shots right first time (timestamps, keyframes, Omni Reference, 21:9)?
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Youri van Hofwegen (2026-10-17) uses one reusable structure: line 1 = style + subject + the hook element; then timestamped beats, each with one action and one camera move ("anything you leave out of a prompt Kling just fills in on its own"); a sound line; a final avoid-list (e.g. warped wings, distorted face; "no rain" stated both in line 1 and the avoid list because storms invite rain). Dialogue goes in quotes inside the timestamp where it is spoken; Kling lip-syncs it. Kling refunds failed generations but charges in full for finished-but-wrong clips, so specificity protects budget; on Kling, 30 s at 720p cost 300 credits vs 540 at 1080p (test at 720p, finalize at 1080...
• articleWhich paid provider models can run inside ComfyUI as built-in "Partner" (API) nodes, and how does that change workflow design?
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. The ComfyUI docs index (captured 2026-10-23) lists 298 built-in Partner node pages: Video 126, Image 101, 3D 45, Audio 16, Text 10. By name prefix the largest groups are Tripo (24), Recraft (20), Bria (19), Kling (19), ByteDance (18, i.e. Seedance/Seedream), Luma (15), Vidu (15), Flux/BFL (12), Wan API (12, incl. Wan 3 reference/image-to-video), Gemini (10), MiniMax (9), PixVerse (9), Meshy (8), ElevenLabs (8), Rodin (7), HeyGen (7), Grok (7), OpenAI (7), Runway (7), Tencent (6), Ideogram (6), plus Magnific, Fish, Topaz, Krea, Reve, LTX, Sync, Beeble, Veo, Claude and Higgsfield entries. Implication: a single ComfyUI graph can mix local open-weight models with hosted clo...
• articleHow do you build a consistent multi-character cast with LoRAs (dataset, captions, training steps) for image and video models, per a July 2026 local workflow?
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Mickmumpitz (2026-07-17) gives five LoRA rules: (1) a made-up trigger word absorbs everything not captioned; (2) vary what should generalize (include wide shots, lighting, poses, clothing), or the character only exists as trained (e.g. close-ups); (3) decide what stays fixed versus editable; (4) captions decide editability: captioned clothes stay changeable but must then be re-prompted; (5) captions are reverse prompts, so match the target model's prompt style (long natural language for Krea 2; structured JSON with bounding boxes for Ideogram 4). Pipeline: an edit-model "character creator" generates a varied synthetic dataset from one or two images (add close-ups of det...
• articleHow can you build a whole AI short shot-by-shot locally with consistent characters and voices (LTX 2.3 + FLUX.2 Klein), and what licences apply?
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Mickmumpitz (2026-05-04): per shot, FLUX.2 Klein generates a start frame from up to 4 references (character face close-up + slightly turned full-body image auto-merged into a sheet; environments/assets too), then LTX 2.3 generates video with audio; an "LTX ID" LoRA plus a voice sample keeps the character voice consistent; each shot group can reference previous shots and color-match the first shot; a shot duplicator and assembler build the edit, and swapping the reference images and voice sample re-casts the whole film without touching prompts (keep character description vague in prompts for this). Paid version adds a 360-degree environment and crop group for consistent ...
• articleWhen T2I-CompBench++ uses multimodal LLMs (GPT-4V, ShareGPT4V, MiniGPT-4) as the judge, how do they compare with specialised metrics, how stable are they, and what failure modes does the paper report?
  Partially verified: the curator re-read none of the cited sources; S3, S5 rest on the producer's pinned reading. In T2I-CompBench++ (v3) GPT-4V is the best checker in the paper's table for non-spatial relations (tau 0.476) and complex prompts (0.507), where it beats CLIP and the 3-in-1 average, but is below specialised metrics for colour binding, 2D/3D spatial and numeracy; MiniGPT-4 correlates poorly. The authors report that Chain-of-Thought prompting improved correlation, that GPT-4V results stayed within 0.032 over five repeats on 50 images (temperature 1.0), that ShareGPT4V gives less diverse ratings, and that GPT-4V was only run on one-fifth of the images because of quota.
• articleAcross China's Measures, Korea's Article 31, California's AI Transparency Act and New York's synthetic-performer ad law, which media, actors and label forms does each text address?
  Scope: synthesis bound to the dated legal-text observations C2-B60-K4-06, -07 and -08; it is not additional evidence. Curator-checked synthesis across China, California and Korea texts: media scope, label form and addressed actors differ as offered; technical specifics are delegated outside the statute in China (national standard) and Korea (decree). The cross-text term mapping is synthesis, not any source's statement.
• articleIn ComfyUI's cache, exactly what happens when a node's IS_CHANGED / fingerprint_inputs returns NaN, raises, or returns a changing value, and does it affect downstream nodes?
  At the pinned commit the fingerprint value is one component of each node's signature, and the signature of a node also contains its ancestors' fingerprints. A NaN, a fresh float NaN produced by an exception, or any value that differs between prompts therefore gives a different cache key, so the node and everything downstream of it is not found in the output cache and runs again. The docs state this for NaN (NaN is unequal to anything); the code and a maintainers' comment agree, but the mechanism is Python equality/hash of nested frozensets, which was read, not run. One edge is inferred from code reading only: a NaN object that is the same object every run could compare equal by identity.
• articleHow does a language tag reach the font's 'locl' glyph substitutions and OpenType language-system tables, and where can that link silently break?
  OpenType 1.9.1 defines 'locl' as localized glyph forms that should always be applied and be driven by language metadata that selects a language-system (LangSys) table; language-system tags are typographic-convention tags, not language codes, and a run uses the specific LangSys or the script default, not both. CSS Fonts 4 requires renderers to derive the OpenType language system from the content language, keeps locl on by default, and provides font-language-override for cases where the font lacks the language; unknown override tags are silently ignored. HarfBuzz documents that shaping input is a (direction, script, language) triple per run with BCP 47 language tags, and activates locl by default. The three layers state their strength differently (OpenType 'may', CSS 'required', HarfBuzz API parameter).
• articleGPT Image 2 prompt structure from fal's guide: five slots, anti-slop rules, change-vs-preserve edits, labelled multi-image roles
  Template: Scene / Subject / Important details / Use case / Constraints, with line breaks; the Constraints slot ("no watermark, no logos, preserve face, preserve layout") is where unbounded prompts drift. Anti-slop rules: replace praise words (stunning, epic, masterpiece) with visual facts (overcast daylight, brushed aluminium, 50mm feel); give style tags concrete visual targets; name the real object (boarding pass, transit kiosk); in edits separate "Change" from "Preserve" and repeat the preserve list each iteration; treat text as typography (quote it, specify font, size, colour, placement; spell hard words letter by letter); one revision per turn. Three modes: generate, edit one image, combine multiple images with each input labelled by role (Image 1: base scene to preserve; Image 2: ...). Consistent with OpenAI's own cookbook (C3-B406-01) and Google's positive-framing rule (C3-B408-01).
• articleFor keeping one person's face identity across storyboard panels, what do the PhotoMaker and InstantID papers and official repositories measure, what limits and compute do they state, and how are their code and checkpoints licensed?
  PhotoMaker (arXiv 2312.04461v1; repo 060b4fc) stacks 1-4 ID photos into one embedding for SDXL; its Table 1 reports PhotoMaker CLIP-I 73.6, DINO 51.5, Face Sim. 61.8 at 10 s on a V100 (IPAdapter has higher Face Sim. 67.1), trained on 8 A100 for two weeks; it handles one person per image and is better at half-length portraits. InstantID (arXiv 2401.07519v2; repo 2145b67) uses one face photo with an InsightFace embedding, a ControlNet-style IdentityNet and an image adapter, trained on 48 H800; the paper has no quantitative table (qualitative comparisons) and the README supports only the largest face. Licence components differ: both codebases are Apache-2.0 (PhotoMaker's LICENSE header mentions third-party components but lists none after the Apache text); InstantID's README says InsightFace face models are non-commercial research only and its released checkpoints are research only, while the Hugging Face card says the project is Apache.
• articleHow does ConsiStory keep one subject consistent across prompts without training, how was that measured, what are its stated failure modes and costs, and can the official code be used commercially?
  ConsiStory (arXiv 2402.03286v3, SIGGRAPH 2024 TOG) shares internal activations between the images of a set so that a subject stays consistent without any optimisation step; the first two prompts act as anchors (paper: two anchor images, 0.5 dropout). The authors evaluate 100 image sets with DreamSim (background removed) and CLIP and a 3,000-response two-alternative user study; the quantitative headline is a plot, not text. Timing: 32 s on an H100 for two anchors plus one new image, about 25x faster than The Chosen One (their estimate of 13 min). Stated limits: it relies on cross-attention localisation (failures below 5% in their observation, fixable by changing the seed), images in a set must share one style, and SDXL biases carry over. The official NVlabs code is under the NVIDIA License, which permits only non-commercial use and only with NVIDIA Processors.
• articleWhat steps, learning rates, class-image counts and VRAM figures do the DreamBooth and textual-inversion papers and the diffusers v0.41.0 docs state, and where do they differ?
  As stated: DreamBooth paper ~1000 iterations, prior-loss weight lambda=1, learning rate 1e-5 (Imagen) or 5e-6 (Stable Diffusion), 3-5 images, ~1000 class samples (fewer possible), about 5 minutes on one A100 for Stable Diffusion. diffusers examples/dreambooth README: 5e-6 with 400 steps (no prior), 5e-6 with 800 steps and 200 class images (prior-preservation), 2e-6 for the 12 GB variant, LoRA DreamBooth 1e-4 and 500 steps, with the text encoder needing at least 24 GB. Textual inversion paper: batch size 4, base lr 0.005 scaled to 0.04, 5,000 steps on 2 V100s, about two hours per concept; diffusers README: 5.0e-04 with --scale_lr and 3000 steps. Hugging Face's own blog experiments report 400 steps at 2e-6 for objects and ~1200 steps at 1e-6 for faces. The values differ by source and model, and are used as stated; none is a universal setting.
• articleWhat do the vendors' terms and help pages say about using user inputs and outputs to train models, by plan or surface, and about restrictions on using outputs to build competing models?
  Training statements vary by surface. OpenAI: consumer content may be used to train with an opt-out; API and business content is not used by default. Google Cloud: will not use Customer Data to train without permission. Adobe: says it does not train on Creative Cloud subscribers' personal content and that partner models do not change that. Midjourney and Ideogram contain no clear 'we train on your content' or 'we do not' sentence in the terms read; Midjourney's privacy policy lists data collected through training its algorithms and grants itself a perpetual licence over inputs and assets; Ideogram's privacy policy lists training the models among uses of information collected. Google Cloud and OpenAI bar using outputs to build similar or competing models, and Midjourney bars using the Services to develop competitive products.
• articleAt the observation dates, do OpenAI, Google Cloud (Imagen), Midjourney, Ideogram and Adobe Firefly say the user owns image outputs and may use them commercially, and does the plan matter?
  OpenAI (consumer terms and business agreement), Ideogram and Google Cloud state in their terms that the vendor does not claim ownership of outputs (OpenAI and Ideogram also assign any rights they hold; Google says Generated Output is Customer Data and it asserts no ownership). Midjourney says the user owns assets but ties ownership to a Pro or Mega plan for companies above one million USD annual revenue, and its help article words the same rule as needing the plan to use images commercially. Ideogram's help FAQ defers commercial use to its terms, while the terms do not restrict it; its plan matters for privacy of images. Adobe's FAQ permits commercial use of non-beta Firefly output but is silent on ownership in the text read. This candidate records texts, not legal conclusions.
• articleHow do open-source film-emulation tools encode manufacturer datasheet curves, and what unit and provenance traps does that introduce?
  In spectral_film_lut at 6307a98, film records store datasheet curves with an explicit exposure base (default 10): the Fuji Eterna 500 record uses base 2 and an x range of about -6.2 to +6.3 (consistent with stops), the Kodak 5219 record base 10 and -3.9 to +0.9 (consistent with log exposure), so mixing records without the base would misplace exposure. Provenance gaps: the Fuji grain curve is copied from Kodak 5207 and the Kodak rms is a computed value. MIT covers the code but says nothing about the datasheet-derived numbers. Fujifilm describes its digital Film Simulation as not merely replicating film.
• articleHow do you turn a traditional blue-pencil sketch plus pencil 'inking' into clean digital line art in Krita, step by step?
  Revoy's 2014 tutorial turns a blue-pencil sketch into clean line art in Krita 2.8.5 (Pentel 0.5 mm blue lead, 300 dpi scan, channel curves so blue drops out, Desaturate Max, Levels, manual cleaning); the page now carries the same 'AI derivation prohibited' footer, in tension with the author's statement that pre-October-2026 material stayed plain CC.
• articleWhat sketch, ink and colour pipelines did Revoy try for episodes 21 and 22 (2017), and why did he keep or drop each?
  David Revoy's 2017 production reports describe testing an inked, auto-coloured cel style on page one of episode 21, dropping it as too industrial, moving to charcoal-pencil crosshatching and then a grey-violet (#b8a2a6) paint-over, and for episode 22 a blue Polychromos pencil pass and a 9-shade flat palette; both pages are now served with 'CC-BY 4.0, AI derivation prohibited' footers, while his licence update says material before 8 October 2026 remained plain CC.
• articleAt diffusers 086bf957 the guide recommends a CPU Generator for reproducibility, and the shipped enable_full_determinism() hard-codes CUBLAS_WORKSPACE_CONFIG=':16:8' while the repository's own test helper switches to ':4096:8' for torch >= 2.10 citing...
  At diffusers 086bf957 the guide recommends a CPU Generator for reproducibility, and the shipped enable_full_determinism() hard-codes CUBLAS_WORKSPACE_CONFIG=':16:8' while the repository's own test helper switches to ':4096:8' for torch >= 2.10 citing CUBLAS_STATUS_NOT_INITIALIZED.
• articleWhat does The Chosen One (iterative consistent-character extraction from a text prompt alone) cost per character, how was it measured, what failure modes do the authors list, and is there official code?
  The Chosen One (arXiv 2311.10093v4, SIGGRAPH 2024) takes only a text prompt, generates a set of images, clusters their embeddings, optimises LoRA weights and two learned text tokens on the most cohesive cluster, and repeats until convergence. The authors report 1-2 iterations, about 13-26 minutes on an A100 (bfloat16), and list the cost as a limitation (about 20 minutes). Consistency is measured as pairwise CLIP image-embedding similarity across prompts and prompt similarity as CLIP score; the main result is a trade-off plot (Fig. 6) plus an AMT user study; the numbers are not in the HTML text. Stated failure modes: identity may not fully converge, a supporting character such as a pet is not kept consistent, spurious attributes can bind to the identity, one character at a time, and simplistic centred scenes. The author project page lists only an 'unofficial' code link. The catalogue already holds this paper as a source entry; this candidate adds the pinned v4 details.
• articleWhat do I need to run StoryDiffusion's consistent-character comic generation for the first time (prompts, GPU, defaults), what does it keep consistent, and what is and is not released or licensed?
  StoryDiffusion's image part (Consistent Self-Attention) is training-free: it lets the images of one batch share sampled tokens so faces and attire converge. Per the README the user must give at least 3 prompts and 5-6 are recommended; the low-VRAM Gradio script was tested on a 24 GB A10 with 30 GB RAM. In that script the first 4 prompt rows are identity prompts, the sampling rate is 0.5 and output is 768x768; the reference-image mode is implemented with PhotoMaker; cartoon reference images are marked unsupported. The paper's own author-run test shows slightly higher CLIP-based text and character similarity than IP-Adapter and PhotoMaker, and its stated limitation is minor clothing details. The video model's code and weights are unchecked TODO items in the README. Code is Apache-2.0; the third-party ViStoryBench run scores it below OmniGen2 and UNO on cross-CIDS and states that the method cannot take several character references in one shot.
• articleWhere have learned models been documented for line work in a drawn-look pipeline (production vs research vs open tools), and what are their inputs, limits and terms?
  Documented production use: Sony Pictures Imageworks trained a scikit-learn GradientBoostingRegressor inside Houdini to predict how artist-drawn ink lines should adjust with camera angle and expression, ~600 frames per character first generation, with artists nudging the prediction (account by Grochola via fxguide, 2019). This model moves hand-drawn lines on 3D animation; it does not extract lines from images. Research/open line extractors differ: Informative Drawings (arXiv 2203.12691v3) maps photographs to line drawings, with a stated dependence on a pretrained depth network; ControlNet 1.1 lineart models use extracted or hand-drawn lines to steer image generation. None of the two research/open sources documents use on 3D character renders.

More