batch-image-generation

# Batch Image Generation ## When to Use **Use this skill when:** - The user needs to produce 5 or more related images from a shared structural template, where some elements change and others stay constant across the series - The user wants to systematically vary one or more dimensions (subject, color, background, mood, style, composition angle, time of day, season) while keeping the rest of the prompt locked - The user is building a product catalog, social media content calendar, character expression sheet, background variation set, or concept exploration grid - The user needs a reproducible, documented workflow so they can regenerate, adjust, or extend the batch later without rebuilding it from memory - The user wants to run parameter sweeps to find optimal settings (CFG scale, stylize, chaos, step count, guidance scale) before committing to a full production run - The user is managing a multi-model batch where different image types are routed to the model best suited for them - The user needs output naming conventions, folder structures, and metadata systems to keep 20-100+ images organized and traceable - The user wants to create "families" of images -- visually cohesive series like seasonal variants, color palette explorations, or mood board expansions **Do NOT use when:** - The user needs a single, highly specific image -- use the model-specific prompting skill instead (Midjourney, DALL-E, Stable Diffusion, or Flux as appropriate) - The user's primary challenge is maintaining a consistent character, face, or personality across images -- use `midjourney-consistency` which covers `--cref`, `--sref`, and identity-locking techniques - The user has a prompt that is producing wrong results and needs diagnostic help -- use `ai-image-prompt-debugging` to fix the underlying prompt before batching it - The user wants to upscale, enhance, or process already-generated images -- use `ai-image-upscaling` or `ai-image-post-processing` - The user wants a curated set of wildly different, unrelated images -- batch generation assumes structural relationships between images in the series; unrelated images don't benefit from a shared template - The user is asking about Stable Diffusion LoRA training or fine-tuning on a custom dataset -- that is a model training task, not a generation batch task --- ## Process ### Step 1: Define the Batch Scope and Classify the Batch Type Before writing a single prompt, identify which batch archetype the user's request falls into. Each has different structural requirements. - **Variation batch:** One subject, multiple treatments. Example: the same product in 6 colorways, or the same character in 8 moods. The subject is always the constant; the treatment is the variable. - **Combination batch (cross-product):** Two or more variables combined exhaustively or selectively. Example: 4 characters x 3 settings = 12 images. Warning -- three variables with 5 options each = 125 images; always recommend curation over full cross-product beyond 30 combinations. - **Series batch:** A sequential set that tells a story or progresses through states. Example: 12 images of a landscape through seasons, or a product launch sequence. Order matters and should be documented. - **Parameter exploration batch:** Same prompt, different model settings. Used to find the sweet spot before committing. Typically 4-16 images. Output is a comparison grid, not a final deliverable. - **Content calendar batch:** A larger production run (20-60 images) organized by theme, week, or campaign phase. Requires sub-template structure -- one master template plus 3-5 sub-templates. Ask the user: - How many total images are needed? - What changes between images? List every varying dimension explicitly. - What must stay identical across all images? (camera angle, lighting style, aspect ratio, artistic style, model version) - Which AI model(s) will be used? (This determines seed behavior, parameter syntax, and workflow tooling) - Is this a draft run for ideation or a production run for publishing? - What is the output resolution and aspect ratio requirement? ### Step 2: Build the Prompt Template with Typed Placeholders A batch prompt template is a reusable structure where every variable element is replaced with a typed placeholder. "Typed" means the placeholder name describes what kind of value fills it, not just a generic `{VAR1}`. **Template structure anatomy:** ``` [SUBJECT_DESCRIPTION], [SUBJECT_ATTRIBUTE], [ENVIRONMENT], [LIGHTING], [CAMERA_DETAILS], [ARTISTIC_STYLE], [QUALITY_BOOSTERS] [MODEL_PARAMETERS] ``` **Placeholder typing convention:** - `{COLOR:matte_white|ocean_blue|terracotta_red}` -- pipe-separated list of valid values - `{SETTING:studio|outdoor|lifestyle}` -- categorical choice - `{MOOD:serene|dramatic|playful}` -- qualitative dimension - `{ANGLE:front-45deg|overhead|three-quarter}` -- specific technical value - `{SEASON:spring|summer|autumn|winter}` -- discrete enumeration **Constant-locking rules for the template:** - Lock the camera angle in text, not just in parameters. "45-degree angle, slightly elevated" is more reliable than assuming the model will default to it. - Lock lighting in specific terms: "soft diffused studio lighting from camera-left" rather than "good lighting." - Lock style era and rendering engine for Stable Diffusion: always specify the model checkpoint name in the workflow, not just in the prompt. - For Midjourney, lock `--v`, `--ar`, `--style`, and `--s` as constants in the template. These must not vary unless the variable is intentional. - Write negative prompts as constants and append them identically to every generated prompt. A negative prompt that differs between images is an uncontrolled variable. **Template validation test:** Substitute two different variable values manually and read both full prompts aloud. If anything other than the intended variable sounds different, the template has a hidden variable. Fix it. ### Step 3: Build the Variation Matrix The variation matrix is the operational heart of the batch plan. It is a complete enumeration of every image to be generated, with its variable values, full prompt, seed assignment, and output filename. **For a single-variable batch (simplest case):** | # | Variable: Color | Full Prompt | Seed | Output Filename | |---|----------------|-------------|------|----------------| | 1 | matte white | [full text] | 1001 | batch_white_1001.png | | 2 | ocean blue | [full text] | 1001 | batch_blue_1001.png | **For a two-variable cross-product batch:** Generate all combinations systematically. For 4 colors x 3 backgrounds = 12 rows. Always sort the matrix so the slower-changing variable is in the outer loop (colors) and the faster-changing variable is in the inner loop (backgrounds). This makes the matrix easier to review and makes regeneration requests easier to handle. **For a curated selection from a large combination space:** - State the full possible combination count explicitly: "4 subjects x 5 moods x 3 styles = 60 possible images." - Present the recommended curated selection (20-25 images) with justification for which combinations were excluded. - Common curation rules: exclude redundant combinations (blue/cold and blue/melancholy overlap), exclude low-value combinations (abstract style defeats product photography purpose), prioritize combinations that serve the stated use case. - Always offer to generate the full matrix if the user wants it, but note the time and cost implications. **Matrix size thresholds:** - 1-15 images: Full cross-product is always fine - 16-30 images: Full cross-product acceptable, recommend reviewing before upscaling - 31-60 images: Recommend curated selection or sub-batching in groups of 15-20 - 60+ images: Mandatory sub-batching; recommend API/automation tooling; draft-then-select workflow required ### Step 4: Define the Seed Management Strategy Seed behavior differs fundamentally between models. Applying the wrong seed strategy produces incoherent batches or wasted generations. **Midjourney seed behavior:** - Seeds in Midjourney influence the noise initialization, which affects overall composition and subject placement, not fine pixel detail. - Same seed + same prompt = highly similar result (but not pixel-identical). The seed is a strong composition anchor. - Same seed + different variable value = similar layout with the variable change applied. This is the core mechanism for comparable batches. - Retrieve a seed by reacting to any Midjourney image with the envelope emoji (✉️) -- the bot DMs you the seed. - Seed ranges in Midjourney: any integer from 0 to 4294967295. For batch work, use memorable anchors: 1000, 1001, 2000, etc. - To generate controlled composition variations within one variable value: use seed, seed+1, seed+2. The increment of 1 produces a similar-but-not-identical composition. - To generate dramatically different compositions: use seeds spaced 1000 apart. **Stable Diffusion / Flux seed behavior:** - In SD and Flux, the seed is deterministic -- same seed + same prompt + same parameters = identical pixel output. This is the most reliable seed system. - For a variation batch: use the same seed across all variable changes to maximize compositional comparability. - For a diversity batch where you want variety: use sequential seeds (1000, 1001, 1002) or use the batch count feature in A1111/ComfyUI which auto-increments seeds. - In ComfyUI, use the "increment" seed mode for batch nodes. In A1111, enable "Increment seed by 1 each batch" in settings. - When using ControlNet poses or depth maps as constants, the seed matters less for composition (ControlNet overrides it) but still affects texture, color, and fine detail. **DALL-E 3 and GPT-image-1 seed behavior:** - No user-exposed seed control. Each generation is stochastic. - Mitigation strategy: Generate 3-4 versions of each prompt variation and select the best per variation. This is the "de facto batch" for DALL-E. - Consistency across a DALL-E batch relies entirely on the constant text quality. Make constants extremely specific. - For critical product/brand work, DALL-E batches require human curation at every step -- do not plan automated pipelines around DALL-E. **Seed documentation rule:** Every batch plan must explicitly state the seed strategy in the output, even if the strategy is "random with manual curation." Undocumented seeds make batches impossible to reproduce. ### Step 5: Define Parameter Variation Strategy and Conduct Pre-Batch Test Never run a full batch on untested parameters. A 5% quality improvement in parameter tuning saves 95% regeneration time on a 30-image batch. **Pre-batch parameter sweep protocol:** 1. Select 2-3 representative prompts from the matrix -- choose ones that span the range of your variables (e.g., the lightest color + simplest background, and the most complex variable combination). 2. Run a parameter sweep on only those 2-3 prompts before committing to the full batch. 3. Document which parameter value wins and lock it as a constant for the full batch. **Key parameter sweeps by model:** *Stable Diffusion / Flux:* - **CFG Scale sweep:** 4, 6, 7.5, 10, 12. Values below 6 often produce washed-out, incoherent results. Values above 10 often introduce artifacts and over-saturation. Sweet spot is typically 6.5-8 for photorealistic subjects, 7-9 for illustrated styles. - **Step count sweep:** 20, 30, 40 steps. Beyond 40 steps, quality improvement is minimal but generation time doubles. 25-30 is the practical sweet spot for most models. - **Sampler sweep:** DPM++ 2M Karras vs. Euler a vs. DPM++ SDE Karras. DPM++ 2M Karras is the most consistent for batch work (deterministic across seeds). Euler a introduces more variation between generations. - **Denoising strength (for img2img batches):** 0.4-0.5 preserves original composition, 0.6-0.75 allows significant variation, above 0.8 is nearly equivalent to txt2img. *Midjourney:* - **Stylize sweep:** `--s 0`, `--s 100`, `--s 250`, `--s 500`. Low stylize values (0-100) produce more literal, less aesthetically "polished" interpretations. Values of 250-500 add Midjourney's trained aesthetic sense. For product photography, use 0-100. For artistic work, 250-750. - **Chaos sweep:** `--chaos 0`, `--chaos 20`, `--chaos 50`. Chaos 0 produces the most predictable, repeatable compositions -- ideal for product batches. Chaos above 30 introduces significant compositional variation -- useful for ideation batches, bad for catalog batches. - **Quality sweep:** `--q 0.5` vs. `--q 1` vs. `--q 2`. For draft/test runs, use `--q 0.5` (50% of standard GPU time). For production, always `--q 1` minimum. `--q 2` rarely produces visible improvement sufficient to justify the 2x generation time. **Parameter budget discipline:** Pick ONE parameter axis to sweep per pre-batch test. Testing CFG, steps, and sampler simultaneously produces uninterpretable results. Vary one parameter at a time, hold all others constant. ### Step 6: Define Output Organization System Without a systematic output organization, batches beyond 20 images become unmanageable. File names like `grid_0023.png` and `upscaled_final_v3.png` destroy any hope of reproducibility. **Naming convention structure:** ``` {batch-id}_{variable1}_{variable2}_{seed}_{version}.{format} ``` - `batch-id`: A short, meaningful project code. `mug-catalog`, `hero-banners-q1`, `char-moods` - Variable values: Shortened but human-readable. `wht` for white, `stu` for studio, `drm` for dramatic - `seed`: The exact seed number used - `version`: `d1` for draft 1, `s1` for selected iteration 1, `up` for upscaled - Format: Always `.png` for AI outputs (lossless; avoids JPEG compression artifacts in intermediate files) **Examples of good file names:** - `mug-catalog_terracotta_studio_1001_d1.png` - `hero-banners_summer_bold_2400_s1.png` - `char-moods_happy_forest_1000_up.png` **Folder structure for batches under 30 images:** ``` {batch-id}/ 00_brief/ # Prompt template, variable list, parameter settings, this plan 01_drafts/ # All first-pass generations, unselected 02_selected/ # Best version per combination, chosen from drafts 03_upscaled/ # Upscaled/enhanced final versions 04_exports/ # Final deliverable formats (resized, color-corrected, platform-specific) metadata.csv # One row per image: filename, full_prompt, seed, parameters, selected_flag, notes ``` **Folder structure for batches 30-100 images (sub-batch organization):** ``` {batch-id}/ 00_brief/ sub-batch-01_{theme}/ 01_drafts/ 02_selected/ sub-batch-02_{theme}/ 01_drafts/ 02_selected/ 03_upscaled_all/ 04_exports/ metadata.csv ``` **Metadata CSV schema (minimum required columns):** ``` image_id, batch_id, variable_1_name, variable_1_value, variable_2_name, variable_2_value, seed, model, model_version, full_prompt, negative_prompt, cfg_scale, steps, sampler, ar, generation_timestamp, selected, upscaled, notes ``` This CSV is the batch's reproducibility record. If someone needs to regenerate image #7 in six months, the CSV contains everything needed. ### Step 7: Define the Workflow Sequence and Compile the Full Plan Assemble the workflow into a concrete, ordered sequence with decision gates. **Standard 5-phase workflow:** Phase 1 -- Template validation (1-4 test images): - Generate 2-3 representative images from the matrix using draft quality settings - Verify that the constant elements are truly constant (same angle, same lighting behavior, same style rendering) - Identify any prompt language that produces inconsistent behavior and fix the template - Do not proceed to full batch until template validation passes Phase 2 -- Parameter sweep (4-16 test images): - Run the parameter sweep on the validated template - Select winning parameters and lock them as constants - Document the winning values in the batch plan Phase 3 -- Full batch generation (all images): - Generate all images in the variation matrix - For Midjourney: generate one image per prompt (not a grid of 4) using `--no` to skip the grid and speed selection -- use the `/imagine` command with `--q 1 --s [locked value]` - For SD/Flux/ComfyUI: use batch processing nodes with the pre-configured seed management - Log every generation to the metadata CSV in real time, not after the fact Phase 4 -- Selection and quality gate: - Review all drafts and flag the best result per variation combination - If any combination has no acceptable result: regenerate that specific combination with a new seed (document the seed change in the CSV) - Target: at least one usable image per matrix row before proceeding to upscaling Phase 5 -- Upscaling and export: - Upscale only selected images, never the full draft set - For Midjourney: use "Upscale (Subtle)" for 2x with minimal change; "Upscale (Creative)" only for artistic images where enhancement is acceptable - For SD: use the SD Hi-Res Fix (2x, denoising 0.4-0.5) or Ultimate SD Upscale in A1111 - Export to platform-required formats and dimensions after upscaling --- ## Output Format ```markdown ## Batch Generation Plan: {Batch Name} ### Overview | Field | Value | |-------|-------| | Batch ID | {short-code} | | Total images | {n} ({breakdown: e.g., 4 colors x 3 backgrounds}) | | Batch type | {variation / combination / series / parameter-exploration / content-calendar} | | Target model | {model name and version} | | Quality tier | {draft / production} | | Estimated generation time | {time per image} x {n images} = {total} (excluding review) | | Estimated fast-pass server cost | {if API-based, estimate token/image cost} | --- ### Prompt Template **Template (copy this, substitute placeholders):** ``` {CONSTANT_SUBJECT_DESCRIPTION}, {COLOR:value1|value2|value3}, {ENVIRONMENT:value1|value2|value3}, {LIGHTING}, {CAMERA_DETAILS}, {STYLE}, {QUALITY_TERMS} {MODEL_PARAMETERS} ``` **Negative prompt (constant, copy identically to all prompts):** ``` {negative prompt text} ``` **Constants (must appear word-for-word in every generated prompt):** - Subject core description: "{exact constant text}" - Lighting: "{exact constant text}" - Camera angle: "{exact constant text}" - Style/rendering: "{exact constant text}" **Variables:** | Variable Name | Placeholder | Value Options | |---------------|-------------|---------------| | {Name} | {PLACEHOLDER} | {val1}, {val2}, {val3} | | {Name} | {PLACEHOLDER} | {val1}, {val2}, {val3} | --- ### Variation Matrix | # | {Variable 1} | {Variable 2} | Seed | Full Prompt | Output Filename |