Choose a workflow from the result you need—not from the longest list of controls. This map tells you what to open, which input matters, what the main control changes, and how to prove the result before adding complexity.
Select the outcome. The route shows the minimum viable workflow; the full comparison remains visible below without JavaScript.
Start from a written description without preserving pixels from an existing image.
Keep some visual structure while changing style, subject, lighting or detail across the whole frame.
Change only a selected region: a hand, face, object, background patch or unwanted element.
Create plausible content beyond one or more original borders without merely stretching the image.
Guide spatial structure more explicitly than a prompt or ordinary img2img can.
Borrow source pixels, structure, visual appearance, or a person’s identity through the method that matches that exact job.
Make a result larger, either by pixel upscaling or by a second diffusion pass that can invent detail.
Produce a transparent PNG or edit layers where alpha is part of the generated result.
Turn a still image into a short motion sequence with Stable Video Diffusion.
Call Original Forge from a script, service, or browser client instead of repeating a form by hand.
Most wrong workflow choices come from treating every image input as the same kind of instruction.
Use Txt2img. The prompt, model and seed establish the first image.
text → pixelsControl prompts, seeds and settings →Use Img2img. Denoising strength controls how much source structure can change.
pixels → variationOpen the img2img guide →Use Inpaint. The mask decides where; masked content and denoising decide how.
pixels + mask → patchOpen the inpainting guide →Use ControlNet. A processed pose, edge, depth or other map guides geometry.
control map → structureThe selected checkpoint, VAE, text encoders, LoRAs, ControlNet model and extension still have to belong to compatible families. Prove the model in a clean generation before diagnosing an advanced workflow.
Review model families and file roles →Combine routes only after each one produces an attributable change on its own.
Text → image
Start from a written description without preserving pixels from an existing image.
Image → variation
Keep some visual structure while changing style, subject, lighting or detail across the whole frame.
Image + mask → local edit
Change only a selected region: a hand, face, object, background patch or unwanted element.
Image edge → larger scene
Create plausible content beyond one or more original borders without merely stretching the image.
Control map → guided image
Guide spatial structure more explicitly than a prompt or ordinary img2img can.
Image → visual guidance
Borrow source pixels, structure, visual appearance, or a person’s identity through the method that matches that exact job.
Image → more pixels
Make a result larger, either by pixel upscaling or by a second diffusion pass that can invent detail.
Prompt / RGBA → RGBA
Produce a transparent PNG or edit layers where alpha is part of the generated result.
Image → short video
Turn a still image into a short motion sequence with Stable Video Diffusion.
JSON request → image
Call Original Forge from a script, service, or browser client instead of repeating a form by hand.
Denoising affects all three, but it cannot replace the correct spatial instruction.
Every source pixel can influence the new image. Use it for a broad transformation.
Learn denoising and resize behavior →Paint or upload the edit region. Include enough nearby context for scale and blending.
Learn mask and crop behavior →Add canvas space in chosen directions and generate continuity across the old boundary.
Plan the canvas and diagnose seams →| Control | Question it answers | Common wrong conclusion |
|---|---|---|
| Denoising strength | How much may diffusion depart from the source or masked content? | It does not choose where the edit happens. |
| Mask mode | Should Forge edit the painted area or everything outside it? | A reversed mask can look like a broken prompt. |
| Masked content | What starting signal fills the edit region: original, fill, latent noise or latent nothing? | “Original” can keep an object you meant to remove. |
| Inpaint area | Should processing use the whole image or concentrate resolution around the mask? | “Only masked” does not mean surrounding context is irrelevant. |
| Mask blur / overlap | How softly should generated and preserved pixels meet? | More blur cannot repair incompatible content by itself. |
Current original Forge exposes Img2img, Inpaint, Inpaint sketch and Inpaint upload through Forge Canvas. It also retains two built-in outpainting scripts. The Mosaic workflow in the researched tutorial is a third-party alternative, not the only Forge route.
One constrains generation; the other increases output dimensions. Both can change detail, but for different reasons.
Edges, pose and depth are not interchangeable. A valid preview map does not prove that the selected model matches it or the base model family.
Run the complete ControlNet workflow →Inspect the current built-in ControlNet ↗Target dimensions alone are not success. Compare edges, faces, typography, repeating textures and the image at 100% zoom.
Open the complete upscale guide →Old demonstrations remain useful for understanding intent, but their tabs and settings are not proof of the current main branch.
The July 2024 Forge announcement links a separate LayerDiffuse repository for transparent image generation and alpha-aware editing. The extension code exposes an always-visible accordion in txt2img and img2img, but its repository still identifies the project as WIP.
A 2024 tutorial shows an integrated SVD tab, a checkpoint in models/svd, fixed image ratios, frame and motion controls, and saved frame sequences. In the inspected current main snapshot, modules/ui.py does not construct an SVD tab and no SVD extension is present in extensions-builtin.
models/svd directory as a clue, not proof of a working module.A pleasing image is not enough to explain why a workflow worked. Keep a baseline that lets you attribute each change.
Load the model and generate once without LoRAs, ControlNet, Hires. fix or third-party extensions.
Keep the model, seed, prompt, dimensions, sampler, scheduler and steps fixed.
Enable one mask, control unit, upscale path or extension. Change its first controlling variable only.
Check what should change and what must stay stable. Inspect the console as well as the image.
Keep the output PNG metadata and a short note for any setting not stored there, including extension commits.
Do not compensate for a wrong route by changing several unrelated controls.
Confirm that the source reached the Img2img canvas, then raise denoising strength in a controlled step. If the seed changed at the same time, restore it before comparing. A prompt alone cannot prove Img2img is using the intended source influence.
Run the controlled img2img baseline →Lower denoising strength and remove unrelated modifiers. If exact geometry matters, move that requirement into a matching ControlNet signal rather than forcing ordinary Img2img to preserve everything.
Check Mask mode. Inpaint masked targets the painted area; Inpaint not masked targets its inverse. Verify the mask preview before changing the prompt or model.
Diagnose the complete inpainting path →Remove the object from the prompt and inspect Masked content. Starting from original preserves its visual signal; fill or a noise-based start can be more appropriate for removal. Include shadows or reflections in the mask when they belong to the object.
Expand one direction first. Check mask blur or overlap, describe the continuation rather than the entire old image, and compare the built-in script behavior at a smaller expansion. Large multi-edge expansion makes the cause harder to isolate.
Run the focused outpainting diagnostic →Confirm the unit is enabled, preview the preprocessor output, match the control model to both the map type and base family, and check weight plus guidance start/end. Compare the same seed with the unit disabled. A visible map with the wrong model is not a valid test.
Diagnose the first failed ControlNet gate →Build and prove one ControlNet unit →Identify which path you used. Extras can introduce model-specific ringing or oversharpening; Hires. fix and SD upscale can invent detail or drift because diffusion runs again. Reduce scale or denoising, then inspect faces, text, edges, seams and repeated textures at 100%.
Diagnose the upscale stage →Separate core UI, built-in scripts and external extensions. Record the previous and current Forge commits, start without third-party extensions, and verify whether the feature exists in current code. Do not reinstall a model merely because an extension tab disappeared.
These answers cover the route choices users most often conflate.
Use Txt2img when no existing pixels need to influence the result. Use Img2img when a source image should contribute layout, color or shape. Denoising strength then controls how far the result can move away from that source.
Ordinary img2img processes the whole source image. Inpainting adds a mask so you can target or protect a region. “Only masked” changes the processing crop, while mask mode decides whether the painted area or its inverse is edited.
Inpainting replaces pixels inside the existing canvas. Outpainting enlarges the canvas and generates beyond an original edge. Both depend on masks and surrounding context, but they solve different geometry.
In the inspected current main snapshot, Img2img can expose the built-in scripts “Poor man’s outpainting” and “Outpainting mk2” in the Script selector. A tutorial that uses Mosaic Outpaint is describing a separate third-party extension.
Use ControlNet when a specific signal such as pose, edges, depth, tiles or another processed map must guide structure. Use ordinary img2img when broad visual similarity is enough. They can be combined, but prove each one separately first.
Hires. fix is a second pass attached to txt2img and can add or change detail. Extras applies a selected pixel upscaler to an existing image. SD upscale is an img2img script that processes an image through diffusion in tiles. Choose by whether new diffusion detail is wanted.
Denoising strength is the first variable to inspect. Lower values generally preserve more source structure; higher values permit larger changes. Keep the seed, prompt, dimensions and model fixed while changing only denoising strength.
No. Some tools display a checkerboard-composited preview alongside the real RGBA output. Download the PNG and inspect its alpha channel or place it over another layer before calling the workflow complete.
SVD tutorials document an older integrated extension. The inspected current main UI does not construct an SVD tab, although an SVD model directory remains in the repository. Treat video as historical until your exact build exposes and completes the workflow.
Record the Forge commit, model and companion files, prompt, negative prompt where used, seed, dimensions, sampler, scheduler, steps, guidance, denoising strength, enabled extensions and every workflow-specific unit. Change one variable per comparison.
The page verifies interface routes in current original Forge code and uses older tutorials only to understand user tasks and historical behavior.
Txt2img, Img2img, Inpaint, Forge Canvas, ControlNet, Extras, Hires. fix, SD upscale and two outpainting scripts.
LayerDiffuse was officially announced in July 2024 as a separate extension and still requires version-specific testing.
See the tested evidence boundary →The SVD tab appears in the 2024 walkthrough but is not constructed by the inspected current main UI code.
Video settings inform likely questions and failure points; their numeric values are not universal defaults.
UI construction · Forge Canvas announcement · ControlNet code · Poor man’s outpainting · Outpainting mk2 · Extras upscale · SD upscale · LayerDiffuse announcement
Code-derived ControlNet architecture; every consequential label was checked against the original repository.