Choose SD 1.5 when the exact checkpoint or a required add-on depends on it. Verify that contract, select sd, prove one clean image at the model’s documented baseline, and add only one matching dependency at a time.
Select the strongest constraint. A real dependency is better evidence than a family reputation or one hardware benchmark.
The publisher identifies an SD 1.x / 1.5 base or fine-tune and documents its file, license, baseline settings, and any required VAE.
Record the model page and file identity, choose sd, then run the clean baseline below.
Stop if the page names SDXL, FLUX, an unsupported derivative, or a required component you do not have.
“SD 1.5” is the architecture boundary. The exact checkpoint page defines what you actually load, how you start, and what you are allowed to do.
Record who published the checkpoint and where it declares SD 1.x / 1.5. A repost, merge, or conversion can have different contents and terms.
Choose the publisher’s inference file. Compare the exact filename, size, and checksum when provided; a plausible name does not prove a complete download.
Preserve the recommended settings with the file. Also record whether the model expects a separate VAE or non-default Clip skip.
Forge’s software license does not grant rights to a checkpoint, fine-tune, LoRA, embedding, ControlNet model, or dataset.
The accessible Stable Diffusion v1.5 Hugging Face page used here explicitly describes itself as an unaffiliated mirror of the deprecated RunwayML repository. We use it only as dated reference documentation for the v1.5 training context and limitations—not as proof that a third-party checkpoint is official.
Read the mirror disclosure and model card ↗VERIFIED Current Original Forge has a dedicated SD 1 engine with CLIP-L, a VAE, 768-wide embedding expectations, and default Clip skip 1; its loader includes that engine in architecture detection. Inspect the engine · Inspect the loader ↗
This run answers one question: can this exact file load, sample, decode, and save on this Forge installation?
Put the completed file in the active models/Stable-diffusion directory. If you use path overrides or shared folders, resolve the effective directory first.
Use the top Checkpoint refresh control or restart Forge. Confirm the selected name matches the file identity you recorded.
sdThe preset reveals SD-oriented controls and starting values. It does not convert a different-family file.
Apply its dimensions, sampler, scheduler, steps, CFG, Clip skip, prompt guidance, and documented VAE state. Use batch size 1.
No LoRA, embedding, ControlNet, Hires fix, face restoration, styles, post-processing, or unrelated extension. Keep only a VAE explicitly required by the checkpoint.
Watch the console. A pass must complete model load, prompt conditioning, sampling, VAE decode, and file save.
Keep the PNG, infotext, seed, checkpoint filename/hash, Forge commit, launch arguments, VAE and Clip skip state, console result, time, and observed memory.
These values come from the inspected Original Forge code. Saved user settings can replace them, and the exact checkpoint page should win when it specifies another value.
Current portrait-oriented start.
Current square start.
Current txt2img start.
Use the model-card value if given.
Current automatic path.
Change only for a documented need.
Automatic/bundled unless specified.
Hires stays off for acceptance.
VERIFIED Commit-bound values from Original Forge modules_forge/main_entry.py at dfdcbab. Inspect the preset code ↗
Family context helps you spot an implausible recipe. It does not override a fine-tune publisher’s tested settings.
The v1.5 reference mirror says fine-tuning used 512×512 images. Forge currently opens sd txt2img at 512×640. For a derivative, use its own tested dimensions; use 512×512 only as a clearly labeled fallback when documentation is absent.
Leave the selector automatic or bundled for a self-contained checkpoint. Choose a separate VAE only when the publisher names it. Wrong VAE state can change color or break final decode.
Current Forge starts the SD control at 1. Some models or embeddings may document 2. Treat that as a compatibility instruction, not a universal quality upgrade.
If a native baseline succeeds, test Hires fix or upscaling as a separate job. That keeps model validity, composition changes, decode problems, and memory pressure distinguishable.
Choose an upscale route →The first level that fails identifies the smallest useful investigation. Never install a whole workflow pack before the base image exists.
One valid image at the documented baseline proves load, prompt encode, sample, decode, and save.
Apply only the exact published requirement and repeat the same seed. Keep the previous result.
Match SD 1.5, use the documented trigger and weight, then compare against the unchanged baseline.
Match both the SD 1.5 family and the task. Preserve the input image and control settings.
Enable the exact extension, Hires pass, post-processor, or script last. A failure here does not invalidate the base checkpoint.
Original Forge is built on the A1111 WebUI codebase, but its backend and Gradio migration changed extension-facing behavior.
The current README identifies Forge as a platform built on Stable Diffusion WebUI and names the exact A1111 upstream commit used as its base.
Read the Original Forge README ↗The Gradio 4 announcement warned that some extensions could break. A later temporary list said some important A1111 extensions had not updated and might need forks or replacements.
Gradio notice · Temporary list ↗If an extension only names AUTOMATIC1111, current Original Forge support remains unknown until the exact repository and revision pass on the exact Forge commit.
Isolate the extension →Record identity. Extension URL and commit; Original Forge commit; Python, Torch, Gradio, OS, GPU, and launch arguments.
Start clean. Confirm the same SD 1.5 checkpoint works without the extension.
Enable one dependency. Restart once, reproduce once, and keep the first console or browser error.
Classify the failure. Installation/import, missing UI, model mismatch, generation, or output behavior require different fixes.
Check each item as it is saved. The receipt is complete only when both model identity and output state are reproducible.
A valid image proves the technical baseline. It does not prove that the model fits your creative goal, license, production volume, or every optional extension.
Keep the last passing receipt and change one variable. Later symptoms are often noise from an earlier failure.
Resolve the active models/Stable-diffusion directory, confirm a finished supported filename, refresh Checkpoint, and distinguish a primary checkpoint from a VAE or add-on.
Use the first console error. Recheck family, file completeness, size/hash, inference-vs-training role, and required components; then retest without extensions or forced precision flags.
Preserve the first error →Restore the checkpoint’s documented VAE state, remove post-processing and extensions, and replay the same seed. The final VAE decode is a separate pipeline stage.
Diagnose decode output →Return to the model’s native baseline at batch 1. If it passes, treat Hires fix or upscaling as a separate workload rather than rewriting the prompt first.
Separate generation from upscale →Match the SD 1.5 family and exact task, verify triggers, inspect skipped or mismatched keys, and compare the unchanged seed with only one component enabled.
Install and validate one LoRA →Recheck the component contract →Prove the base without it, record both commits, enable only that extension, restart, and separate install/import, UI, model, generation, and output errors.
Run extension isolation →The checkpoint baseline has passed; the later stage is the failing workload. Record the stage and measured memory, then reduce one driver such as batch, canvas, ControlNet, or Hires workload.
Route the allocation stage →These answers keep family support, settings, add-ons, extensions, performance, and licensing as separate claims.
Yes. The inspected Original Forge source has a dedicated StableDiffusion engine matched to the SD15 architecture, and the loader includes it among recognized model families. This proves support in the checked code snapshot, not that every SD 1.5 conversion or add-on is valid.
Download the exact inference checkpoint from its publisher, keep the model page and license, place the finished file in the active models/Stable-diffusion folder, refresh Checkpoint, select it, choose the sd preset, and generate one base-only image.
The default primary-checkpoint folder is models/Stable-diffusion. Path arguments or shared-folder configuration can change the active location, so resolve the effective path when a correct file does not appear.
Use sd. In the inspected code it shows the VAE and Clip skip controls and supplies SD-oriented starting values. A preset changes the interface; it does not identify, download, convert, or repair a checkpoint.
Use the exact checkpoint publisher’s documented resolution. The reference v1.5 model was fine-tuned at 512×512, while current Forge opens the sd txt2img preset at 512×640. Both facts are starting context, not a rule that every derivative shares one canvas.
That is the current portrait-oriented txt2img default in the inspected UI code. It can be replaced by saved settings, and the exact checkpoint documentation should override it for the acceptance run.
Use the model card’s tested baseline first. If it gives no dimensions, 512×512 is the documented training context for the mirrored v1.5 reference, but it is still an explicit fallback assumption—not a universal fine-tune recipe. Test larger aspect ratios only after one valid baseline.
Follow the exact checkpoint documentation. Forge’s current sd preset starts with Euler a, Automatic scheduling, and CFG 7, but those UI defaults are not a universal recommendation for every SD 1.5 fine-tune.
The inspected Forge sd preset exposes Clip skip and starts at 1. Use 2 only when the exact model or embedding documentation requires it. Changing Clip skip alters prompt conditioning; it is not a generic quality switch.
Only when the checkpoint publisher requires a particular VAE or the package intentionally omits one. Start with the documented bundled or automatic state. A random VAE can change color and decode behavior without fixing a family or file error.
Confirm the active checkpoint directory, completed filename, supported extension, and refresh state. Current checkpoint discovery includes .ckpt, .safetensors, and .gguf, but VAE-named files are excluded from the primary checkpoint inventory and belong in their own role.
Read the first console error. Verify the publisher, family, complete download, size or hash when provided, and whether the file is an inference checkpoint rather than a training-only component. Retest without extensions and forced precision experiments.
Return to the model’s documented VAE state, turn off extensions and add-ons, and rerun the same seed. A preview followed by a broken final image points toward decode or post-processing; a failure before sampling belongs to another stage.
Restore the checkpoint’s native baseline before changing prompts. A much larger first-pass canvas changes the denoising problem and memory load. Prove the native image, then test Hires fix or an upscale workflow as a separate stage.
No. A shared .safetensors container does not make tensor shapes or text conditioning compatible. Use an add-on whose publisher explicitly identifies the same SD 1.x / 1.5 base family.
Confirm the base family, trigger syntax, weight, and console messages. Compare the same prompt and seed with only that add-on changed. Discovery in the UI proves the file was indexed, not that it applied successfully.
Do not assume so. Match the control model to both its declared control task and base architecture. Use the dedicated ControlNet workflow to validate one component at a time.
Some may, but SD 1.5 checkpoint support does not establish extension compatibility. Original Forge’s dated Gradio 4 notice warned that extensions could break, and its later temporary list said some important A1111 extensions needed forks or replacements. Test the exact extension and revision.
This page makes no universal speed or memory promise. Checkpoint implementation, precision, dimensions, batch, VAE, add-ons, GPU, driver, Forge commit, and measurement method all affect results. Measure your complete job locally.
Not when a required checkpoint, LoRA, embedding, control model, or production workflow depends on its ecosystem. If no dependency exists, compare current families against the actual job instead of choosing only by release age.
Forge’s AGPL-3.0 software license is not permission for a separate model or add-on. Read the exact checkpoint, derivative, LoRA, embedding, and dataset terms for the intended use. This guide is not legal advice.
Dated tutorials helped us identify user vocabulary and questions. They do not supply universal settings, performance promises, or current extension compatibility.
Repository identity, SD engine, architecture detection, checkpoint inventory, UI preset, and Clip skip option at dfdcbab.
The latent-diffusion paper supplies architecture background. The v1.5 page is explicitly an unaffiliated mirror and is cited with that limitation.
O13 and O20 establish that the 2024 Gradio migration could affect extensions and that forks/replacements were sometimes needed; they are not a 2026 compatibility list.
V01 and V27 supplied checkpoint, resolution, workflow, and performance-search vocabulary. V27 was a narrow creator test and was not generalized into speed or VRAM claims.
Tested environment: static site build and browser QA; product facts checked against Original Forge source at dfdcbab (26 Jun 2025). Author: Forge Field Guide editorial team. Reviewer: Forge Field Guide technical review. Updated: 1 Sep 2026.