Image generation

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

• articleSDXL-Turbo: 1 step, guidance 0, 512x512 and the img2img steps x strength rule
  SDXL-Turbo: 1 step with guidance_scale=0.0 (no CFG, no negative prompt), 512x512 preferred. For img2img keep steps x strength >= 1 (2 steps at 0.5 strength = 1 step). Licence: non-commercial sai-nc-community.
• articleQwen-Image: documented resolutions and diffusers settings
  The Qwen-Image model card quick start uses a fixed resolution per aspect ratio (1328x1328 for 1:1, 1664x928 for 16:9), 50 steps, true_cfg_scale 4.0 and a language-specific quality suffix on the prompt.
• articleMidjourney Style Explorer: Try Style, Copy and saved style codes
  To reuse Midjourney style codes, Try Style reruns the current or latest prompt; Copy adds the --sref value to the Imagine bar. Liking a style saves it without changing personalization profiles.
• articleThe Explore page allows users to browse community-created images and videos using filters like 'For You', 'Random', 'Hot', and 'Top'.
  The cited source states: The Explore page allows users to browse community-created images and videos using filters like 'For You', 'Random', 'Hot', and 'Top'.
• articleBest practices for Style Reference include keeping text prompts simple and avoiding conflicting style descriptors.
  The cited source states: Best practices for Style Reference include keeping text prompts simple and avoiding conflicting style descriptors.
• articleThe captured documentation says an uploaded image cannot be converted into a style code. --sref random applies a random style code; to use an uploaded image as a Style Reference, f
  The cited source states: The captured documentation says an uploaded image cannot be converted into a style code. --sref random applies a random style code; to use an uploaded image as a Style Reference, follow the image-reference workflow instead.
• articleStyle Reference Codes can be applied using `--sref <code>` or `--sref random` from Midjourney's internal library.
  The cited source states: Style Reference Codes can be applied using `--sref <code>` or `--sref random` from Midjourney's internal library.
• articleStyle weight (`--sw`) ranges from 0 to 1000, defaulting to 100, and has varying impact on V7 depending on usage with style codes vs. images.
  The cited source states: Style weight (`--sw`) ranges from 0 to 1000, defaulting to 100, and has varying impact on V7 depending on usage with style codes vs. images.
• articleThe `--sref` parameter with an image URL is used for Style Reference in Discord prompts.
  The cited source states: The `--sref` parameter with an image URL is used for Style Reference in Discord prompts.
• articleOn the Midjourney website, click the image icon in the Imagine bar to upload or select an image, then drag it from the uploads library into the Style Reference section. In Discord,
  The cited source states: On the Midjourney website, click the image icon in the Imagine bar to upload or select an image, then drag it from the uploads library into the Style Reference section. In Discord, append --sref and an online image URL to the prompt.
• articleStyle References are compatible with Midjourney versions 6 and later.
  The cited source states: Style References are compatible with Midjourney versions 6 and later.
• articleMidjourney Style Reference captures visual vibe without copying objects or people.
  The cited source states: Midjourney Style Reference captures visual vibe without copying objects or people.
• articleHow easily can invisible image watermarks be removed, and which watermark types resist regeneration or adversarial attacks?
  Corrected 2026-10-24 - see C3-B1132-01; the original statements below are kept for history and are superseded where the correction says so. Corrected 2026-10-23 - see C3-B681-04; the original statements below are kept for history and are superseded where the correction says so. Two benchmark-style papers test removal. WAVES (5 watermarks, 26 attacks) finds no attack works on all watermarks: regeneration/diffusion attacks are most effective on Stable Signature (average detection performance 0.001 under Regen-Diff), Tree-Ring is most vulnerable to adversarial attacks, and StegaStamp is most robust but visibly degrades images. Zhao et al. report over 99 percent removal for DwtDctSvd, RivaGAN and SSL by regeneration, with StegaStamp needing heavy noise. The two papers disagree on Tree-Ring under regeneration, probably because of different strengths and metrics, and both use attackers without detector access.
• articleHow can an image or layout agent check a composition against grids, and which checks are documented vs only proposed?
  Corrected 2026-10-24 - see C3-B1068-01; the original statements below are kept for history and are superseded where the correction says so. Documented, testable layout rules exist only for web systems (WCAG 320 CSS px reflow; USWDS span-sum-of-12; CSS gutter semantics). Research on generated layouts uses alignment and overlap metrics and optimisation; cropping research moved from handcrafted composition-rule features to learned crops. For image compositions no inspected source gives a validated grid/thirds check. Any such check (e.g. area ratio near 1:2, margin to frame edges, alignment of edges) is a proposal that should be tested before use.
• articleWhich structural control (depth, segmentation, edges, blur, normals) should I weight to preserve which property, according to the ablations that report it?
  Corrected 2026-10-24 - see C3-B1087-02; the original statements below are kept for history and are superseded where the correction says so. In Cosmos-Transfer1's TransferBench results, blur-visual and edge controls give the highest adherence and lowest diversity, while depth and segmentation allow more variation; raising a modality's weight in a region raised its alignment metric there (Pearson 0.93 and -0.92 in the authors' test). RealMaster's ablation found edges preserved faces better than depth for propagation. EPE found G-buffer inputs matter for realism. Normals appear as a control only in vendor text (Reallusion) and EPE's G-buffers, not in the Cosmos Transfer1 README or VACE's task table.
• articleWhat is the reliable way to get correct typography on an AI graphic: generate without text and set real type afterwards, or extract text from a flat image (Ideogram Layerize, Recraft no_text)?
  Corrected 2026-10-24 - see C3-B955-08; the original statements below are kept for history and are superseded where the correction says so. Problem. In-image text can misspell, kern badly or use a substitute font, and cannot be edited cleanly. Conditions. Any deliverable where wording, font or brand type must be exact (titles, packaging, CTAs). Method. (1) Generate a text-free base. Recraft V3 has controls.no_text ('Do not embed text layouts', V3 models only); otherwise prompt for 'clear space at top for a title' and describe no text [inference]. (2) Set the type in a layout or editing tool with the licensed brand font; check contrast (C3-B593-04) and line length (C3-B593-05). (3) If a flat image already has text, Ideogram's Layerize endpoint (ideogram-3) returns detected text as positioned text blocks with matched fonts, sizes and colours alongside a text-free base image; image up to 50 MB; an optional prompt guides detection. 'Matched fonts' means similar, not necessarily identical...
• articleIf I POST exactly the same API-format prompt twice, does ComfyUI generate a new image? What does control-after-generate have to do with it?
  Corrected 2026-10-24 - see C3-B879-01; the original statements below are kept for history and are superseded where the correction says so. Problem. Scripts that loop POST /prompt with a fixed seed and expect new files, or that expect the seed to advance. Conditions. ComfyUI v0.42.5, default cache mode, same server process, no IS_CHANGED returning NaN on the output chain. Method. 'control_after_generate' is only a flag in the node's input options (e.g. KSampler seed in nodes.py); a search of server.py, execution.py and cli_args.py finds no backend handling, so seed advancement is performed by the frontend before it builds the prompt. An API-format prompt therefore always carries a concrete seed. When the whole signature chain is cached, execution.execute() finds cached output for each node, calls emit_cached_output (re-emitting the stored UI result) and never runs the node function; at the end, cached nodes with HAS_INTERMEDIATE_OUTPUT also re-emit. The websocket reports the node lis...
• articleHow do the vendors' docs describe mask-based editing and inpainting (mask format, polarity, fidelity), and what do they say the mask does not guarantee?
  Corrected 2026-10-24 - see C3-B977-02; the original statements below are kept for history and are superseded where the correction says so. Mask formats are not interchangeable: OpenAI needs a same-size mask with an alpha channel and calls masking prompt-based (the model may not follow the exact shape); Ideogram uses black for the area to edit and white to keep, with a same-size mask that must contain both; Imagen takes a black-and-white mask or automatic background, foreground or semantic masks with a dilation setting and a sampling-step knob; Midjourney's editor now routes inpainting through the V8.X Edit Model. Ideogram 4.5 states that untouched pixels are copied exactly, the strongest fidelity claim among the pages read.
• articleHow do the FLUX, Imagen, Gemini and OpenAI docs tell me to write text I want rendered in an image (quoting, length, font description), and what limits do they admit?
  Corrected 2026-10-24 - see C3-B955-06; the original statements below are kept for history and are superseded where the correction says so. Problem. Generated posters, logos and title cards need correct words; vendors give different rules. Conditions. Vendor docs fetched 23 Oct 2026; model names change quickly. Method (documented). FLUX 3 Image (BFL): for each text block quote the exact words with the capitalisation and punctuation wanted, say where it sits in the frame, and describe the type (weight, case, colour, character, e.g. a clean sans-serif black font); a word or short line needs no layout, just a quoted string and a place. Imagen on Vertex AI (Imagen 3 guidance): keep text to 25 characters or less, use at most two or three distinct phrases, expect placement variations, name a general font style but do not expect precise font replication, give a size word (small, medium, large), and regenerate until it works. Gemini image docs: be clear about the text, the font style (descri...
• articleWhat reference-based tools do vendors document for keeping a character, subject or style consistent across images, what limits do they state, and where do the docs disagree?
  Corrected 2026-10-24 - see C3-B976-03; the original statements below are kept for history and are superseded where the correction says so. Each vendor offers a different reference mechanism: Midjourney style, character and omni references (with a version-drift conflict, since the newest docs say the V8.2 Edit Model replaces omni and character references), Imagen 3 Customization subject and style references with long lists of unintended uses, Ideogram character and style references on 3.0 and reference images on 4.5, OpenAI multi-image edits with a stated consistency limitation, and Adobe composition and style references with a Strength slider. Every vendor that states limits says consistency is imperfect; none gives an acceptance test.
• articleWhich output formats, compression levels and transparent-background settings do the OpenAI image generation and edit endpoints document, and which combinations are valid?
  Corrected 2026-10-24 - see C3-B955-05; the original statements below are kept for history and are superseded where the correction says so. Problem. Choosing format, compression and background so a generated asset keeps alpha and loads fast. Conditions. GPT image models on the Image API (generations and edits). Method. output_format is png (default), jpeg or webp. output_compression is an integer 0-100 that applies only to jpeg and webp; the openapi schema gives it a default of 100 on edits, and the generation schema constrains it with minimum 0 and maximum 100. The guide says jpeg is faster than png when latency matters. background is transparent, opaque or auto (default); transparent requires png or webp output (jpeg has no alpha). The spec says gpt-image-2.5-sunburst and -flare support opaque and transparent backgrounds, while for gpt-image-2 and gpt-image-2-2026-04-21 transparent support is 'in preview'. Failure modes. transparent with jpeg is outside the documented combination. ...
• articleHow do the OpenAI and Google image-editing guides tell me to keep identity, camera angle and lighting unchanged in an edit (change-only-X prompts, input_fidelity, text masks)?
  Corrected 2026-10-24 - see C3-B955-01; the original statements below are kept for history and are superseded where the correction says so. Problem. Fixing one thing in a keyframe (wardrobe, time of day, object) without changing the face, framing or style. Conditions. gpt-image-1.5 edits (OpenAI Cookbook) and Gemini image models (Nano Banana 2/Pro, Google Cloud guide). Method. (1) State what changes and what stays: 'change only X' + 'keep everything else the same', and repeat the preserve list on every iteration to reduce drift. (2) Lock invariants explicitly: preserve identity, geometry, layout, camera angle, object placement. (3) For heavier edits that must keep a face, pass input_fidelity="high" in images.edit (cookbook examples for lighting/weather change, person-in-scene, compositing); the guide says higher input fidelity helps maintain likeness during larger scene edits. (4) Lighting or weather re-stage: change only lighting direction/quality, shadows, atmosphere, precipitation...
• articleWhat do the vendors' own docs promise about seed reproducibility, and which conditions (watermark setting, session, mode, order) break it?
  Corrected 2026-10-24 - see C3-B976-02; the original statements below are kept for history and are superseded where the correction says so. The five vendors state five different strengths of seed promise. Imagen on Vertex AI is the strongest but conditional: same seed gives the same images, only when addWatermark is false, and image order is not guaranteed. Adobe's Firefly API says the same seed, prompt and presets give the same image. Midjourney calls seeds the initial noise, says they are 99% identical in V8.X, not reliable across sessions and not for Turbo mode. Ideogram says only that seed is for reproducible generation and returns the seed used. The OpenAI image guide has no seed parameter in the text read.
• articleFor a first image request, which size/aspect, count and prompt-rewriting controls do Midjourney, OpenAI, Imagen on Vertex AI, Ideogram and Adobe document, and which defaults change my prompt before generation?
  Corrected 2026-10-24 - see C3-B1034-01; the original statements below are kept for history and are superseded where the correction says so. Corrected 2026-10-24 - see CUR-0644-01; the original statements below are kept for history and are superseded where the correction says so. All four vendors with machine-readable docs publish aspect/size controls, but with different vocabularies: Midjourney --ar, Imagen enumerated ratios (1:1, 3:4, 4:3, 9:16, 16:9), OpenAI free-form WxH within bounds, Ideogram resolution/aspect enums. Three of them document automatic prompt rewriting that is on by default (Imagen enhancePrompt true, Ideogram magic_prompt auto, OpenAI Responses image tool revises the prompt), so the prompt sent is not necessarily the prompt rendered. Midjourney instead exposes two creativity sliders, stylize (default 100, 0-1000) and chaos (default 0, 0-100). Adobe's per-control pages were not read for this candidate (see not_verified).
• articleWhich aspect ratios and resolution tiers (512px, 1K, 2K, 4K) does the Gemini API image doc list per Nano Banana model, and what is the default for generation versus edits?
  Problem. Choosing aspect_ratio and image_size for a Gemini image model. Conditions. Gemini API response_format {type: image} with aspect_ratio and image_size fields. Method. By default the model matches the output size to the input image, otherwise it makes 1:1 squares; set aspect_ratio and image_size (1K default, 2K, 4K, and 512px on 3.1 Flash Image only) under response_format. The doc has per-model tables of exact pixels. Nano Banana 2.1 and 3.1 Flash Image share 14 ratios including extreme 1:4, 1:8, 4:1 and 8:1 (for example 1:1 is 1024x1024 at 1K, 2048x2048 at 2K, 4096x4096 at 4K; 16:9 is 1376x768 at 1K). 3.1 Pro Image lists 10 ratios (no 1:4, 1:8, 4:1, 8:1), and Gemini 2.5 Flash Image lists the same 10 at a single resolution (1:1 is 1024x1024). Gemini 3.1 Flash Lite Image supports only 1K. Failure modes. Pixel sizes are fixed per ratio, not free-form (unlike OpenAI's custom sizes) so an arbitrary size is not possible; the 2K token count differs by model, not covered here. The do...
• articleOpenAI's docs say edit images must be under 50MB but the mask under 4MB: which source says which, and are they actually in conflict?
  Problem. Two numbers circulate for OpenAI edit masks: 50 MB and 4 MB. Conditions. images/edits with GPT image models. Method. The guide's 'Mask requirements' section reads 'The image to edit and mask must be of the same format and size (less than 50MB in size)'. The openapi.yaml CreateImageEditRequest says each input image (png, webp or jpg, up to 16 images) is less than 50MB, but the mask field says 'Must be a valid PNG file, less than 4MB, and have the same dimensions as image'; transparent (alpha zero) areas mark where to edit and the mask applies to the first image. Reading: the guide's 50MB can be read as applying to image and mask together as a loose statement, the schema separates them, so the strict reading is image under 50 MB, mask PNG under 4 MB. The deprecated variations request also says 'less than 4MB' for a square PNG, so a 4MB figure appears in two places and may be remembered for the wrong endpoint. Failure modes. A large mask PNG (for example a 4K full-colour mask)...
• articleWhat output sizes and aspect ratios can I request from OpenAI's image API, and which are only for the newer GPT Image models (arbitrary WIDTHxHEIGHT within limits, 3840 max edge) versus the three fixed sizes?
  Problem. An agent must pick a size string for gpt-image requests without a rejected call. Conditions. Image API generations and edits; model matters. Method. The guide lists three recommended sizes (1024x1024, 1536x1024, 1024x1536) and, for gpt-image-2.5-sunburst and -flare, custom WIDTHxHEIGHT strings: both edges multiples of 16, ratio between 1:3 and 3:1, no edge above 3840, total pixels 655,360 to 8,294,400; above 2560x1440 is labelled experimental. The openapi.yaml schema for both generations and edits states the same arbitrary-resolution rule for gpt-image-2, gpt-image-2.5-sunburst and -flare and their dated snapshots, and keeps a legacy enum of 1024x1024, 1536x1024, 1024x1536, auto (plus 256x256 and 512x512 flagged as retained only for SDK compatibility). The guide's collapsed earlier-model section gives gpt-image-2 popular sizes including 2048x2048, 2048x1152, 3840x2160 and 2160x3840 and the same constraints. Failure modes. Sending a free-form size to older models such as gpt...
• articleWhat did IPTC's own 2019 test find about metadata on Twitter, Pexels and Pixabay saves?
  Problem. Historical evidence of upload stripping. Conditions. 2019 test with reference image. Method. Upload image with many IPTC fields, then check displayed, saved and downloaded copies. Failure modes. SavedAs copies lose metadata. Limitations. Dated, tests IPTC and Exif only. Version/date. Spring 2019; fetched 2026-10-24 UTC.
• articleDoes Cloudflare Images keep embedded metadata when transforming by URL?
  Problem. Whether a CDN transform keeps provenance tags. Conditions. Cloudflare Images transform via URL, metadata option. Method. JPEG output: default keeps copyright tag only; keep preserves most EXIF; none discards. WebP or PNG: all discarded. Failure modes. Converting to WebP drops metadata. Limitations. IPTC/XMP/C2PA undocumented here. Version/date. Fetched 2026-10-24 UTC.
• articleWhat does IPTC say commonly strips embedded photo metadata on a website?
  Problem. Embedded metadata not visible on published images. Conditions. Image passes through a DAM, CMS or CDN. Method. Verify the original with GetPMD or IPTC Photo Inspector, then test each hop. Failure modes. WordPress/Drupal, DAM or CDN stripping. Limitations. Names categories only. Version/date. Fetched 2026-10-24 UTC.
• articleWhich wins in Google Images when IPTC embedded metadata and structured data disagree, and why does embedding help?
  Problem. Which creator/licence data Google shows. Conditions. Image has both IPTC and schema.org structured data. Method. Embed IPTC once per image; add structured data per page if you need to override; conflicts resolve to structured data. Failure modes. Stale structured data hides correct embedded values. Limitations. No statement about AI-origin fields. Version/date. Fetched 2026-10-24 UTC.
• articleHow does Google Photos present IPTC-based AI info, and when does the section disappear?
  Problem. Whether a Google Photos 'How this was made' panel is reliable. Conditions. Photo with IPTC metadata only. Method. Open info panel and scroll to 'How this was made', or use Lens > About this image. Failure modes. Panel missing when no AI edit is recorded or metadata nonstandard; metadata may be edited or removed. Limitations. Absence proves nothing. Version/date. Fetched 2026-10-24 UTC.
• articleWhat does Google's 'About this image' show about how an image was made, and how reliable is it?
  Problem. What can a user or agent learn from Google's image panel? Conditions. Image Viewer > About > Learn more about an image. Method. Read the 'How the image was made or edited' section (camera capture versus AI generation, 'Info from' names the confirmer) and validate with other sources as Google advises. Failure modes. Treating the section as proof; absent section does not mean image is real. Limitations. Inputs not enumerated. Version/date. Page undated; fetched 2026-10-24 UTC.
• articleWhich IPTC Digital Source Type terms are composites or retired, and which should an agent stop writing?
  Problem. A reader meets composite or retired values in files. Conditions. Any file with Digital Source Type. Method. composite makes no claim about AI; compositeCapture says all elements are real captures; compositeSynthetic (C3-B788-01) says at least one is generative. digitalArt (retired 2024-09-17) maps to digitalCreation; softwareImage retired 2022-06-14. Reuse under CC BY 4.0 with attribution. Failure modes. Reading composite as 'AI-free' or as 'AI'. Limitations. Retired terms may still appear in old files. Version/date. Fetched 2026-10-24 UTC.
• articleWhat do the IPTC terms algorithmicMedia, digitalCreation, algorithmicallyEnhanced and dataDrivenMedia mean, and are they AI terms?
  Problem. Agents confuse non-generative algorithmic terms with AI terms. Conditions. Reading or writing Digital Source Type. Method. Treat algorithmicMedia (no training data, e.g. maths formula), digitalCreation (human, non-generative tools), algorithmicallyEnhanced (sharpening, noise reduction) and dataDrivenMedia as non-generative; use C3-B788-01 terms for generative content. Failure modes. Mapping diffusion output to algorithmicMedia. Limitations. Edge cases such as AI denoise or upscale are not covered by the definitions. Version/date. algorithmicMedia and dataDrivenMedia created 2022-06-14; digitalCreation 2024-09-17; fetched 2026-10-24 UTC.
• articleWhich three IPTC Digital Source Type terms should an agent use for AI-generated, AI-edited and mixed-with-AI images, and what are their exact definitions?
  Problem. Agents need exact IPTC wording to tag or interpret AI image provenance. Conditions. Image or video metadata field Digital Source Type (URI under cv.iptc.org/newscodes/digitalsourcetype/). Method. Use the term whose definition matches the workflow: trainedAlgorithmicMedia (display name 'Created using Generative AI', modified 2024-10-23), compositeWithTrainedAlgorithmicMedia ('Edited using Generative AI', created 2023-07-26) and compositeSynthetic ('Composite including generative AI elements'). Failure modes. Choosing a human-readable name instead of the URI; using trainedAlgorithmicMedia for inpainting edits. Limitations. The scheme page defines terms only; the mapping of real tool outputs is inference. Version/date. Page fetched 2026-10-24 UTC; entry modified dates 2022-2024.
• articleDoes the MOSS-Image tutorial's licence link actually cover the weights? The linked GitHub README says only pipeline code is Apache 2.0; the Hub model card says weights and code are.
  Problem. The tutorial's licence line links to the GitHub repository, whose README licence section covers only the standalone pipeline code. Conditions. Recording licence for the MOSS-Image Base or SFT checkpoints used in ComfyUI. Method. Cite the Hub model card instead: the OpenMOSS-Team MOSS-Image-Base card states the model weights and bundled inference code are Apache 2.0 (the SFT card also carries apache-2.0 front matter via the Hub API and a LICENSE file in its file list). Failure modes. Citing the GitHub README for weight rights leaves a gap. Limitations. The Comfy-Org int8 and gptq re-quantisations carry their own apache-2.0 label; I did not check the Qwen3.5 9B text-encoder terms. Version/date. Fetched 2026-10-24.
• articleIs Mage-Flow safe for commercial use because the tutorial says MIT? The MIT licence file is real, but the upstream README also says the models are released for research purposes only and not for product or service deployment.
  Problem. The ComfyUI Mage-Flow tutorial states only 'License: MIT'. Conditions. Any use of Mage-Flow outputs or weights in a paid or deployed workflow. Method. Read both: the per-model licence table lists Mage-Flow as MIT, and a separate README paragraph says the models are released for research purposes only and are not intended for product or service deployment. Record both statements; treat the intended-use sentence as a risk flag to resolve with the licensor, because the sources do not say which governs. Failure modes. Citing MIT alone overstates what the vendor's own README permits as intended use. Limitations. I did not read a separate model-card on the microsoft HF org (the Hub API returned an auth error for microsoft/Mage-Flow), so a weights-specific term may exist elsewhere. Version/date. Fetched 2026-10-24.
• articleCan I paint over an AI-generated base texture in Blender, and what UV limits break it?
  Problem. AI texturing (Paint3D, Hunyuan3D-Paint) leaves seams, baked shadows or wrong detail to fix by hand. Conditions. Mesh with a UV map and the AI texture loaded as an Image Texture node. Method. Texture Paint mode paints into the UV-mapped image; save the image (or pack it) or round-trip via an external editor (GIMP, Krita): edit, re-save, then reload in the Image Editor. Failure modes. Overlapping UVs are not generally supported, which matters when AI UV atlases reuse islands: a stroke across faces sharing texels paints both. Limitations. Texture slot, brush and projection details not read in depth. Inference: re-bake or re-unwrap overlapping islands first. Version and date. Blender 5.2 LTS manual (updated 2026-10-23).
• articleDoes Google review the product images and prompts I use in Product Studio, and what do reviewers see? Quality reviewers may see inputs, outputs and instructions.
  Problem. Pre-release or confidential product photos could be exposed to human reviewers when put through a free generative tool. Conditions. Product Studio in Merchant Center or the Shopify Google & YouTube app. Method. Per the page, a team of quality reviewers may review inputs and new images or videos; reviewers see the original offer images, the generated images or videos and the instructions given. The page says only the information needed to assess quality is shared and confidential business information is protected. To use the tool a seller must accept the Terms of Service and, by country, Product Studio Additional Terms. Failure modes. Uploading unreleased products or embargoed designs. Limitations. The page does not give retention periods, opt-out or reviewer location; terms were not read in this round. Version and date. Undated page fetched 2026-10-24.
• articleDoes Google's Merchant Center Product Studio video page say AI labels are required, and are they exempt from the overlay and watermark ban? Labels are allowed for certain jurisdictions; no compliance guarantee.
  Problem. Sellers adding visible AI disclosure to a product video or image worry it breaks the no-text/no-watermark rule. Conditions. Generated or AI-edited ad assets in Merchant Center, Google Ads and related products. Method. (1) The page states that AI regulations in the European Union, India and New York require disclosures or labels on ads with certain AI-generated or edited assets. (2) Add the label inside the creative or use the AI label setting in Google Ads, Display & Video 360, Campaign Manager 360, Merchant Center and Ads Editor. (3) The page says such labels will not violate the policies on text overlays and watermarks. (4) It also says using the setting does not guarantee compliance, so take legal advice. Failure modes. Assuming the setting satisfies a given law; relying on in-creative text without checking the image-spec rules in C3-B783-02. Limitations. The statement is on the video page, so whether the same exemption governs still images in image_link is not stated th...
• articleWhere do AI-generated lifestyle scenes belong in a Google Shopping feed, and what are the lifestyle_image_link limits? Optional attribute, up to 5 images, 600 x 600 minimum, no borders, AI tag required.
  Problem. A generated scene (product in a room, on a model) is not a clean packshot, so it may fit the lifestyle attribute better than the main image. Conditions. lifestyle_image_link is optional, used for in-context images shown in free listings. Method. (1) Submit up to 5 lifestyle URLs, comma-separated. (2) Use at least 600 x 600 pixels and a valid aspect ratio. (3) Apply the same overlay ban: no calls to action, price, shipping, warranty text, watermarks, logos or barcodes. (4) Do not add a border or padding to hit an aspect ratio. (5) Keep the AI-origin metadata in the file. Failure modes. A generated scene that has a border, padding, retailer logo or baked-in text is rejected and not shown. Limitations. The page's aspect-ratio text reads 'between 2:0 and 2:3', which looks like a typo, so the true allowed range is unverified. That lifestyle suits generated scenes is my inference; the page does not mention AI scenes except in the metadata line. Version and date. Undated page fetc...
• articleHow should a seller prepare input images and publish Product Studio results so the generated image becomes the main or an additional Merchant Center image? One product, centred on white; multiple selected outputs go in as additional images only.
  Problem. Generated scenes may overwrite or not replace the main product image depending on how they are saved. Conditions. Product Studio, change-background (Generate image) flow in Merchant Center. Method. (1) Input: pick one starting image with a single main product, centred with most of the canvas white; use the 'reduce' button if the product touches the edge; avoid images of people using the product and avoid hands or props. (2) Scene description: state image style, product type, placement, surroundings and background (page example: cologne bottle on marble with petals and tropical leaves). (3) The tool returns four outputs; review each at full size. (4) To set an output as the main image, select exactly one image and tick the 'main' option in the 'Add this image to Merchant Center?' dialog. (5) If several images are selected, all are added as additional images, never as the main image. Failure modes. Selecting two or more outputs silently routes them to additional images; the p...
• articleWhat does Google say Product Studio cannot do or should not be used for (hands, wall art, light fixtures, regulated goods, availability)? Experimental; works best on simple packaged goods; unsupported categories listed.
  Problem. Before building a workflow on an AI background-replacement tool, a seller needs the stated scope limits. Conditions. Product Studio inside Merchant Center (and the Google & YouTube app on Shopify). Method. Check each product against the stated limits: it works best for consumer packaged goods with the product front and centre; images with hands, wall art and light fixtures are unsupported; image or video creation is unsupported for regulated goods (alcohol, drugs, pharmaceuticals, gambling-related products, weapons, tobacco, fireworks, health and medical devices); available to all Merchant Center users except those in RU and FR (Shopify app: except RU); the Generate Video feature is limited to Australia, Canada, India, Japan, the UK and the US; descriptions work best in ten listed languages. Failure modes. The page warns the feature 'may generate unexpected images' (so every output needs human inspection against the real product, an inference about practice). Limitations. '...
• articleWhat are Google Merchant Center's main-image technical limits and overlay bans, and which of them bite hardest on AI-generated product images? Minimum 500 x 500 from 31 Jan 2027; no added text, logos or watermarks.
  Problem. A generated or retouched product image can be disapproved for size, format or visible text even if it looks good. Conditions. Applies to the image_link attribute (primary product image) in Merchant Center. Method. (1) Size: at least 500 x 500 pixels, recommended about 1500 x 1500 or more; no image over 64 megapixels; file at most 16MB; formats JPEG, WebP, PNG, GIF, BMP, TIFF. (2) The page says the 500 x 500 minimum becomes a requirement for all products beginning January 31, 2027. (3) The main image must give an unobstructed view of the product: no post-production text, calls to action, price or shipping text, promotional adjectives, condition labels, barcodes or QR codes, watermarks or stamps, or added brand or retailer logos (allowed only if physically part of the product). (4) No placeholder or generic graphic. (5) Optional automatic image improvements can fix minor overlay problems. Failure modes. Image models often render legible-looking text, logos or badges into pack...
• articleDoes Google Merchant Center require AI-generated product images to carry IPTC DigitalSourceType metadata, and which values does it list? Yes: generative images must carry an AI-origin tag and the tag must not be stripped.
  Problem. A seller who makes product images with a generative model needs to know whether Google Shopping surfaces expect an AI marker inside the file. Conditions. Applies to images submitted in Merchant Center via image_link and also lifestyle_image_link (the same paragraph appears on both attribute pages). Method. (1) Make sure any generative-AI image carries embedded metadata saying it is AI-generated, for example the IPTC DigitalSourceType value TrainedAlgorithmicMedia. (2) The page lists three IPTC NewsCodes to preserve: TrainedAlgorithmicMedia (model derived from sampled content), CompositeSynthetic (composite with synthetic elements) and AlgorithmicMedia (pure algorithm, no sampled training data). (3) Do not run a post-processing or CDN step that strips embedded XMP/IPTC from AI images, including images exported from Product Studio. (4) Check an exported file with an IPTC-aware reader before upload. Failure modes. Image optimisers, resizers and screenshots commonly drop metada...
• articleGiven only these public pages, how should an agent build and caveat a cost estimate for an AI-assisted shot list?
  Corrected 2026-10-24 - see CUR-0618-01; the original statements below are kept for history and are superseded where the correction says so. Proposed procedure (synthesis): list shots with duration, resolution and audio; map each to a price cell in one unit family; round durations up to the vendor allowed values; multiply by an assumed take count (no source states a retake rate); add stills at per-image prices; check against spend and concurrency caps; report a range with capture dates. Worked example, 12 shots x 6 s x 3 takes = 216 generated seconds at 720p: Veo 3.1 Lite $10.80, Gemini Omni Flash about $21.89, Veo 3.1 Fast $21.60, Runway gen4.5 $25.92, Veo 3.1 Standard $86.40 (arithmetic in work/estimate.py; assumptions, not measured). The sources limit this: failed-generation billing, retakes, taxes and account discounts are mostly unstated, and some pages disagree with themselves.
• articleWhen I add LPIPS or CLIPScore to a generation regression test, which input conventions, versions and meaning limits must I respect?
  Problem. SSIM is brittle to harmless shifts; teams add learned metrics but misuse them. Conditions. LPIPS reference implementation (PerceptualSimilarity repo, commit 082bb24f, LPIPS version 0.1) and the CLIPScore paper (arXiv 2104.08718v3). Method. LPIPS: lpips.LPIPS(net='alex') is the default and best as a forward metric; net='vgg' is closer to a traditional perceptual loss for optimisation; inputs must be RGB tensors Nx3xHxW scaled to [-1, +1]; output is a distance (lower means more similar). The README asks publications to state the version used. CLIPScore is defined as CLIP-S(c, v) = w * max(cos(c, v), 0) with w = 2.5, reference-free, between a candidate caption and image embedding; the authors computed it with a ViT-B/32 backbone. Practical regression design (inference): record metric name, backbone, version and input scaling in the test config; use LPIPS for image-to-image change (baseline vs new render) with a tolerance calibrated on your own repeat-run noise, and treat CLIPS...
• articleHow does ComfyUI's own repository run image-quality regression tests (baseline directory, SSIM threshold, matching by prompt metadata), and what can I copy or must I fix for my own workflows?
  Problem. Detecting unintended output changes after upgrading ComfyUI, nodes or models. Conditions. ComfyUI v0.42.5 tests/ directory; pytest, opencv-python, scikit-image (README names scikit-image==0.21.0). Method. Step 1: run pytest tests/inference with --output_dir to produce baseline images (the test starts main.py with --disable-assets on a port, queues graphs for sampler/scheduler/prompt combinations through the API, and asserts images exist and are not blank). Step 2: change code. Step 3: run inference again plus tests/compare/test_quality.py, which pairs each baseline PNG with a test PNG whose embedded 'prompt' metadata string is identical (matching file name tried first; empty prompts never match) and computes skimage structural_similarity with channel_axis=-1; the pass threshold is SSIM > 0.95; it writes PASS/FAIL rows, a difference image and a grid image. Failure modes. Metadata match requires identical prompt JSON, so any workflow edit orphans its baseline; only SSIM is im...
• articleWhat does ComfyUI-layerdiffuse provide for transparent and layered generation, and which constraints (SD1.5-only joint modes, batch sizes, diffusers pin) apply?
  Problem. Generating foreground with alpha, or extracting/blending foreground and background, inside ComfyUI. Conditions. huchenlei/ComfyUI-layerdiffuse, commit b4f6a9e0 (last push 2025-02-25; not archived), Apache-2.0 code, requirements diffusers>=0.29.0 and opencv-python. Method. Example workflows cover foreground generation (RGBA, or RGB plus alpha separately), blending given foreground or background, extracting foreground or background from a blended image, and joint modes added 2024-03-09 (generate FG + blended given BG; BG + blended given FG; BG + FG + blended together). Joint modes need batch size 2N (or 3N for all three) and currently SD1.5 only. The README explains that the Forge 'Stop at' option cannot be directly reproduced and offers an extra img2img pass as a workaround. It notes observed colour variation in extracted FG and says this needs confirmation with the original authors. Failure modes. Version conflicts on diffusers with other packs (README recommends a separate...

More