Find where the image turns wrong

Save the failed receipt, restart once, choose the correct sd/xl/flux preset, use one documented model package, batch 1 and a fixed seed, set low bits to Automatic, remove custom launch flags, and disable Hires. fix plus every add-on. Run the same baseline twice, then restore one stage at a time.

The live preview may use Approx NN, Approx cheap, or TAESD instead of the final Full VAE. Compare the saved final file first. If only the preview is wrong, changing the model package or final-decoder precision is premature.

That shape moves attention to Full VAE decode and later postprocessing. Disable every later stage, verify the exact VAE or FLUX autoencoder, and compare Full decode with one isolated TAESD test. It narrows the stage but does not prove one universal cause.

On the inspected commit, the warning occurs when Forge converts the decoded, clamped image to uint8. Non-finite values survived to that conversion. The message does not reveal whether they began in the sampler, VAE, add-on, or retained state.

No. A wrong or missing VAE is one proven case, especially for split FLUX packages, but black output is also reported with sampling, checkpoint switches, MPS environments, ControlNet, extensions, risky flags, and postprocessing. Isolate the first bad stage.

That exact wording can appear in older or different WebUI code. In the inspected original Forge commit, test_for_nans returns immediately. Capture the repository identity and commit before following advice written for another code path.

No. The flag does not repair non-finite values. On the inspected commit, the relevant test function already returns without checking, so hiding a check cannot turn invalid image data into a valid result.

Not as a first step. It broadly changes precision and can hide the stage that matters. First prove whether sampling or Full VAE decode is the boundary; then use one code- and hardware-justified comparison and record its cost.

Only as an isolated decode-stage comparison after a clean baseline proves sampling passes. Record the startup VAE dtype and remove the flag if it does not change the same fixed-seed result.

TAESD is a different approximate decoder. The result shows the latent can be rendered through that alternate path, which focuses diagnosis on Full decode or its component. TAESD quality and resolution behavior can differ, so it is not the final repair by itself.

Record the VAE / Text Encoder selections, startup VAE dtype, the console “Model selected” object, and PNG Info. For single-file models, also check whether the publisher says the VAE is baked in; for split packages, record the explicit autoencoder filename.

First verify the flux UI preset and one complete documented package: diffusion model, CLIP-L, one intended T5 variant, and the correct autoencoder when the package is split. Then test Automatic low bits, Queue, CPU, batch 1, no add-ons, and the same seed twice.

The maintainer specifically identified the official ae.safetensors file in issue #1278 for that user’s split FLUX setup. A bundled checkpoint may have a different contract, so follow the publisher’s package documentation rather than adding files blindly.

No. The multiselect lists discoverable additional modules; it is not a compatibility picker. Select only the exact companion set documented for the checkpoint, and keep every filename in the receipt.

A preset mismatch changes visible controls and family defaults, and issue comments report stale XL controls with a FLUX checkpoint after PNG Info transfer. Verify preset and checkpoint together before generation; treat the report as community evidence, not a guaranteed cause.

That comparison proves the environment can generate something, not that every SDXL file is valid. Test one complete SDXL checkpoint with the xl preset, its documented VAE state, preset base dimensions, batch 1, Automatic low bits, and no add-ons.

Reproduce a minimal A → B → A sequence after a clean restart. Keep batch 1, no Hires. fix, no LoRAs, and fixed seeds. An open issue reports switch-specific black/distorted output, but its loaded workload does not establish a universal root cause.

A black image alone is insufficient. Compare publisher size or checksum, capture load warnings and missing keys, and test another complete documented package in the same baseline. Redownload only from the publisher or trusted original source.

Prove the same generation without ControlNet, then add one compatible unit with one preprocessor, one family-matched control model, and fixed resize/input settings. The full ControlNet workflow page handles pairing; this page only isolates the first failing boundary.

Compare control-map and output aspect ratios, resize mode, Pixel Perfect state, and final dimensions. Disable Hires. fix and other units. If the base output is normal, restore the one control unit without changing the canvas.

Test the base checkpoint first, then add one exact family-compatible LoRA at a documented weight. If it fails, repeat with third-party extensions disabled and record whether the first bad stage is conditioning, sampling, or final postprocessing.

Hires. fix adds later resize, VAE and sampling work. Save the passing base image, use batch 1 and a smaller target, then record whether failure begins before the second sampler or only at final decode. Route an allocation traceback to the OOM page.

Yes. Fully or partly transparent pixels can appear black on a dark background. Open the saved PNG in an alpha-aware editor and place it over a bright color. Do not change precision until you inspect the alpha channel.

Open the exact file elsewhere and compare alpha and PNG Info. If both viewers show black, the failure happened by final conversion/save; if only Forge’s gallery differs, preserve the file and troubleshoot the viewer boundary.

Use the correct preset, one documented component set, base dimensions, fixed seed, and no conditioning add-ons. Noise that never resolves can indicate a different model-contract or sampling problem than a final decode that turns an otherwise plausible preview black.

Reduce to batch 1 and run the exact seeds individually. A mixed batch is useful evidence of intermittent numeric or state behavior, but it does not identify VAE, sampler, hardware, or extension without single-seed reproduction.

Keep the last passing commit and first failing commit, reproduce on a clean extension-free install, and use the same model files and receipt. Rollback can restore work, but a useful regression report still needs the smallest changed commit range.

Include original Forge identity and commit, install route, GPU/driver/Torch, preset, checkpoint hash, every VAE/encoder, launch flags, complete parameters, add-ons, first bad stage, exact warning, saved output, last passing receipt, clean-install result, and a minimal repeat sequence.