Control txt2img instead of chasing lucky seeds
Start with the exact model recipe, make one clean image, freeze its resolved seed and complete receipt, then change one prompt or generation control. That turns “looks different” into a comparison you can explain and repeat.
VERIFIED Original Forge only · source-audited at dfdcbab · no GPU image claim is made.
A good txt2img run answers one visible question
Forge gives access to many controls. The model and the user job decide which ones matter.
Forge provides the controls
Current code exposes prompt fields, sampler and scheduler, steps, dimensions, guidance, batches, seed and Hires. fix. Presets change several displayed defaults by model family.
This guide provides the test method
The contact-sheet protocol and prompt block pattern are our editorial method. They are not official Forge requirements or universal quality settings.
Need only the first successful output? Use Generate your first image. Need source pixels to guide the result? Use Img2img. Need a second-pass enlargement? Use the upscale decision guide.
Know which question each control answers
A control’s available range is not a recommendation to use the extremes.
Prompt
What the image should visibly contain.
- Use it to
- Change one requirement at a time after a baseline works.
- Boundary
- Text can be altered by selected Styles or add-ons.
Negative prompt
Conditioning for what should be discouraged when the active guidance path uses it.
- Use it to
- Start empty unless the model’s documentation requires terms; add only a defect you can name.
- Boundary
- At CFG Scale 1, current UI makes this field non-interactive.
Sampling method
The sampling algorithm selected for the denoising path.
- Use it to
- Keep it fixed while testing steps or scheduler.
- Boundary
- Names and supported combinations can change with the build.
Schedule type
The schedule paired with the sampler.
- Use it to
- Record it separately, even when set to Automatic.
- Boundary
- Automatic resolves through the selected sampler’s configured default.
Sampling steps
The number requested from the current sampler path.
- Use it to
- Find a sufficient interval for the exact model recipe.
- Boundary
- Current UI exposes 1–150 and initially constructs the slider at 20.
Width × Height
The generated canvas and latent shape.
- Use it to
- Use the model family’s documented size; change ratio deliberately.
- Boundary
- Current sliders expose 64–2048 in steps of 8; a preset may replace displayed values.
CFG / Distilled CFG
Model-family-dependent guidance controls.
- Use it to
- Use the model author’s recipe, then compare one guidance control.
- Boundary
- Forge presets deliberately expose different values and visibility for SD, XL and Flux.
Batch count / size
Repeated iterations versus images handled in one batch.
- Use it to
- Keep both at 1 for diagnosis; expand only after one result works.
- Boundary
- Total requested images are count × size; Batch size is displayed up to 8.
Seed
The noise identity used to start a generation.
- Use it to
- Use -1 to search; copy the resolved integer before comparing changes.
- Boundary
- The same seed is not a cross-machine or cross-build pixel identity guarantee.
Hires. fix
A second, larger diffusion pass attached to txt2img.
- Use it to
- Leave it off while proving the base image.
- Boundary
- It has its own denoising, scale, steps, guidance and optional model/sampler controls.
Preset evidence: the inspected main_entry.py supplies different txt2img dimensions, guidance visibility/values, sampler and scheduler choices for sd, xl and flux. One copied “best settings” card cannot represent all three.
Make one replayable image before optimizing
The baseline removes enough uncertainty that the next change has a name.
- 01
Open Txt2img and match the UI preset
Select sd, xl or flux to match the loaded model family, then select the intended checkpoint and required VAE / text encoders. Do not infer compatibility from a filename.
- 02
Start without optional pipelines
Turn off Hires. fix, ControlNet, LoRAs, Styles and third-party scripts for the clean baseline. A model-required component is not optional; record it.
- 03
Write one observable prompt
Describe a subject, action or pose, setting and composition in plain visible terms. Add lighting or medium only when the task needs them.
- 04
Use the model’s documented recipe
Take sampler, scheduler, steps, guidance and native size from the exact model source. Forge exposes controls; it does not make one recipe correct for every model.
- 05
Generate one image
Keep Batch count 1 and Batch size 1. Watch the console for model loading, warnings, OOM or fallback instead of judging only the gallery.
- 06
Freeze the resolved seed
If Seed was -1, use Reuse seed or copy the integer shown in the output metadata. Generate again without changing anything to test the local baseline.
- 07
Save the full receipt
Keep prompt, negative prompt, model hashes, VAE / encoders, seed, sampler, scheduler, steps, guidance, dimensions, preset, add-ons and Forge commit.
- 08
Change one variable
Run a small contact sheet plan with the receipt locked. Stop when the named requirement is answered; do not tune unrelated controls into the same comparison.
One valid image exists; its resolved numeric seed and complete effective setup are recorded; repeating the unchanged local job produces a comparable result; optional workflows are still off.
Write visible requirements, then remove contradictions
A prompt is easier to debug when each phrase has an observable job.
Subject
What must be present? Name count, defining attributes and relationships only when visible.
one ceramic teapotAction / pose
What is the subject doing, and where is it facing or placed?
resting near the table edgeSetting + composition
Where is the scene and how should the frame read?
wood table, eye-level close shotLight + medium
Add the visible treatment only after the scene itself is correct.
soft window light, product photoone ceramic teapot resting near the edge of a wooden table, eye-level close shot, soft window light, product photograph
Change only soft window light to hard afternoon side light. Keep the seed and every other receipt field fixed.
ASSUMPTION This block order is a readable testing convention, not official syntax. If the model source requires trigger words or a different prompt format, that documented contract comes first.
Plan one-variable contact sheets
Select the variable. The desk states what changes, what stays locked and what makes the comparison invalid.
Prompt
- Change
- One visible requirement: subject, action, setting, composition, light, medium or finish.
- Hold fixed
- Checkpoint, VAE / encoders, seed, negative prompt, sampler, scheduler, steps, guidance, dimensions, batch and add-ons.
- Observe
- Did the named visual property change without losing requirements that were already correct?
- Invalid when
- A style preset, LoRA, checkpoint or random seed also changed.
Baseline · Variant A · Variant B
Labels are a plan, not generated images.
Documented model recipe
First prompt alternative
A second prompt alternative
Record the setup the image actually used
A seed without its model and sampling path is not a reproducible record.
Metadata helps, but inspect it. Styles, extension state, precision choices or companion-file hashes may not be complete in every saved record. Add a note for anything consequential that the PNG does not preserve.
Return to the first changed contract
Do not compensate for a model, seed or memory problem with a longer prompt.
The image ignores part of my prompt
Reduce the job to one observable requirement, confirm the correct model and trigger syntax, and check whether Styles or add-ons rewrote conditioning. Then compare guidance or prompt wording—one at a time.
The result changes when I reuse the seed
Compare the complete receipt, not only prompt and seed. Check model/VAE hashes, scheduler, dimensions, guidance, preset, precision, extensions and whether a random seed was actually resolved.
Every random result is poor
Stop seed hunting. Prove the model at its documented sampler, scheduler, steps and native size with optional tools off. A broken model stack does not become healthy through more seeds.
More steps do not improve the image
Keep the sampler, scheduler and seed fixed, then retain the lowest tested interval that meets the visible requirement. More steps are a cost, not a universal quality guarantee.
CFG makes the image harsh or strange
Restore the documented family baseline. Change only the guidance control that the loaded model uses and inspect contrast, saturation and artifacts. Do not import SD 1.5 values into Flux.
The negative prompt cannot be edited
Check CFG Scale. In the inspected UI code, the negative prompt becomes non-interactive when CFG Scale equals 1. This is a UI behavior, not proof that every model interprets negatives the same way.
Changing dimensions ruins composition
Return to the model’s documented size, lock the seed, then change one aspect ratio. A larger canvas also changes memory demand and can cross into offload or OOM behavior.
Batch output seeds look wrong
Inspect each image’s infotext and your exact commit. Historical original-Forge issues reported random-seed and seed-label problems; treat an unresolved mismatch as COMMUNITY-REPORTED and reproduce on a clean install.
Generation is slow or runs out of memory
Return to Batch size 1, disable Hires. fix and reduce the canvas to the proven baseline. Then use the dedicated memory controls and OOM guides instead of changing the prompt.
Fix Forge out-of-memory errors →The first image is black, blank or corrupted
Stop tuning txt2img. Prove checkpoint, VAE/encoders, precision and the clean extension state through the black-or-blank diagnostic.
Diagnose black or blank images →Answers for the decisions users actually type
These questions were audited as long-tail and conversational variants of the page’s single txt2img intent.
What is txt2img in Stable Diffusion WebUI Forge?
Txt2img generates an image from text conditioning without using source-image pixels. The loaded checkpoint and companion files, prompt, seed, sampler path, guidance and canvas define the effective job.
Where is txt2img in Forge?
Open the Txt2img top-level tab, then use its Generation panel. If you only need the first successful image, follow the first-image guide before optimizing parameters.
How do I use txt2img in Forge?
Match the UI preset and model, disable optional pipelines, write one observable prompt, use that model’s documented baseline, generate one image, freeze the resolved seed, save the receipt and change one variable.
What is a good Forge txt2img prompt?
A useful test prompt names visible facts: subject, action or pose, setting and composition, with lighting or medium only when needed. This is our editorial starting pattern, not an official Forge syntax.
Should I copy long prompts from an image gallery?
Use them as dated clues, not proof. Gallery metadata may omit model files, extensions, Styles or settings; start with the smallest prompt that expresses your task and add one requirement at a time.
What should I put in the negative prompt?
Only a model-documented baseline or a defect you can identify in your own outputs. An inherited list can conflict with the model or obscure which term changed the result.
Why is the negative prompt disabled in Forge?
At the inspected commit, changing CFG Scale to 1 makes the negative-prompt component non-interactive. Restore the documented CFG behavior for your model before assuming the field is broken.
What does seed -1 mean in Forge?
The seed UI is constructed with -1, and processing resolves a random seed before generation. Copy the resulting integer from the output before running a controlled comparison.
How do I reuse the last seed in Forge?
Use the Reuse seed control beside Seed; its code describes it as mainly useful after a randomized generation. Verify the numeric seed in the resulting infotext.
Will the same seed make the exact same image?
Only as a local test when the entire effective setup also matches. Different commits, model bytes, VAE, sampler/scheduler, dimensions, precision, add-ons or hardware paths can change the result.
Do seeds increase across a Forge batch?
Current processing constructs per-image seeds from the resolved base seed and image position when variation strength is zero. Historical issues reported wrong labels or random-seed behavior, so verify each image’s infotext on your build.
What is the best sampler for Forge txt2img?
There is no verified universal best sampler. Use the exact model author’s recommendation, keep scheduler and steps fixed, then compare one alternative on a frozen seed.
What does Schedule type do in Forge?
It selects the scheduler paired with the sampler. Automatic resolves through the sampler configuration, so record the field and effective build instead of treating it as meaningless.
How many sampling steps should I use?
Use the model’s documented baseline, then compare separated step counts with the same seed, sampler and scheduler. Keep the lowest interval that meets the named visible requirement.
Does more sampling steps mean better quality?
No universal rule is supported. Extra steps can add time without a useful visible change, and the outcome depends on the model, sampler and scheduler.
What CFG Scale should I use in Forge?
Use the model-family recipe. The inspected Forge presets themselves differ: SD, XL and Flux do not share one displayed guidance configuration.
What is Distilled CFG Scale?
It is a separate guidance control exposed by Forge and made visible in the inspected Flux preset. Record it only when the loaded model workflow uses it; do not substitute it blindly for CFG Scale.
What width and height should I use?
Use the model’s documented native size and aspect ratio. Current controls allow 64–2048 in steps of 8, but availability is not a quality or memory guarantee.
What is the difference between Batch count and Batch size?
Batch count repeats iterations; Batch size requests multiple images in one batch. Total requested images are their product, but memory behavior differs, so prove Batch size 1 first.
How do I generate several variations safely?
Prove one image, then increase only Batch count while keeping the receipt stable. Inspect every result’s resolved seed and do not hide failures.
Should I enable Hires. fix before finding a good seed?
No. Prove the base canvas first. Hires. fix adds a second diffusion pass and more controls, so it makes prompt and seed diagnosis harder and can increase memory demand.
How do I compare two prompts fairly?
Use the same model files, seed, sampler, scheduler, steps, guidance, dimensions, batch and add-ons. Change only the prompt text and write down the exact difference.
Why does a Forge preset change my txt2img settings?
The current Forge preset handler updates dimensions, guidance visibility/values, sampler and scheduler for sd, xl and flux. Saved UI defaults can also influence what appears.
Where are txt2img settings saved with the image?
Forge builds an infotext containing prompt and generation parameters such as steps, sampler, schedule, guidance, seed, size and model details. Saving behavior and metadata format depend on Settings.
Is txt2img the same as img2img?
No. Txt2img has no source-image pixels. Img2img encodes an existing image and adds denoising, so source composition and colors can influence the output.
Why is my txt2img image black or blank?
Treat it as a model/runtime failure before a prompting failure. Check checkpoint, VAE/encoders, precision and extensions with the dedicated black-or-blank guide.
Current code outranks dated recipes
Videos reveal real user questions. They do not establish current labels or universal parameter values.
Official README identifies original Forge, its A1111 base, installation/status boundaries and warns that UI/functionality can change.
Inspect source ↗Current Forge UI code defines sd/xl/flux preset-specific dimensions, guidance, sampler and scheduler values.
Inspect source ↗Current core UI code verifies Width, Height, batch controls, guidance fields, Hires. fix and pasteable generation fields.
Inspect source ↗Current sampler UI verifies separate Sampling method, Schedule type and Sampling steps controls, including the 1–150 step range.
Inspect source ↗Current seed UI verifies -1, Random seed and Reuse seed controls.
Inspect source ↗Current processing verifies random seed resolution, per-image seed construction, batch iterations and infotext inputs.
Inspect source ↗Fresh official-code check verifies sampler–scheduler normalization and the meaning of Automatic.
Inspect source ↗A dated full walkthrough demonstrates prompt, model-recipe checks, random/reused seeds, Styles, Hires. fix and output metadata.
Inspect source ↗A dated full beginner walkthrough covers prompt fields, samplers, schedulers, steps, dimensions, guidance, batches and seed iteration.
Inspect source ↗A short dated overview explains the original Forge UI context and install/model folder flow; its old performance claims are not reused here.
Inspect source ↗An open 2024 issue reports random seed behavior after a PNG-info-to-img2img path; it is not generalized to clean current txt2img.
Inspect source ↗A closed 2024 issue reports incorrect seed labels across batch count/size on a named old commit.
Inspect source ↗Not GPU-run here. This page was checked against original Forge source, full local transcripts and browser behavior. It does not claim that one sampler, step count, CFG or prompt style wins on every model and device.
- Author
- Forge Field Guide editorial team
- Technical review
- Original Forge code + complete catalog transcripts
- Inspected environment
- Original repo · commit dfdcbab · source audit only
- Updated
- 2 September 2026