Choose the exact variant first, identify whether its VAE and text encoders are bundled or separate, then prove one base image before tuning memory or adding a LoRA.
A successful load requires the model identity, complete layout, correct controls, accepted terms, and enough headroom for the actual workload.
Record the repository and commit. A fork tutorial can expose different loaders, controls, defaults, and supported formats.
Record dev, schnell, or the fine-tune’s declared upstream. The variant determines guidance, step, and license decisions.
Separate model identity from FP16, FP8, NF4, GGUF, packaged, and split-file labels.
Save the exact model page and current license before the file loses its source context.
Define acceptable load success, time, memory, stability, and output before changing controls.
Text encoders, VAE, precision, resolution, batch, LoRA patching, system RAM, shared memory, driver, and Forge settings all contribute to the measured path.
Select the strongest description. The output tells you the minimum load contract and the first reason to stop.
Variant: dev · Layout: packagedDo not rank a dev NF4 file against schnell as though only one property changed.
Guidance-distilled 12B model. Current Forge applies Distilled CFG to recognized dev models. The BFL page currently gates files behind acceptance of its dev non-commercial license.
Separate 12B distilled variant. Current Forge ignores Distilled CFG when it recognizes schnell. The BFL card currently lists Apache-2.0 and a 1–4-step reference.
These labels affect storage, precision, loading, computation, and package layout. They do not replace the variant, fine-tune identity, license, or benchmark conditions. Choose and test GGUF vs NF4 → Install and troubleshoot a FLUX LoRA →
This is the launch-level checklist. The complete path resolver remains in the dedicated model-folder guide.
models/Stable-diffusionPackaged checkpoint, raw model, or GGUF. It appears under Checkpoint.
models/VAEThe documented FLUX VAE, commonly ae.safetensors, selected in VAE / Text Encoder.
models/text_encoderThe documented CLIP-L and T5-XXL files, selected together in VAE / Text Encoder.
The inspected scanner accepts .ckpt, .pt, .bin, .safetensors, and .gguf for additional-module inventories. Discovery does not prove that the selected state dict is the expected FLUX component.
The first run answers one question: can this exact FLUX set load, condition, sample, and decode on this machine?
Record the commit and launch arguments. Watch the console from model selection through final decode.
fluxThe preset exposes FLUX controls and starts txt2img at 896 × 1152, CFG 1, Distilled CFG 3.5, Euler, and Simple in the inspected code. These are UI defaults—not a contract for every fine-tune.
Packaged: choose the checkpoint and leave modules empty unless documented. Split: select the model, VAE, CLIP-L, and T5 as one set.
Begin with Diffusion in Low Bits: Automatic, Swap Method: Queue, and Swap Location: CPU. Preserve the preset GPU Weights value for the first observation.
Dev and schnell do not share one step/guidance contract. Use the exact distribution’s documented dimensions and steps; keep batch size 1.
No LoRA, ControlNet, Hires fix, refiner, face restoration, styles, or unrelated extensions in the acceptance run.
Separate cold model load plus first generation from the next warm generation. A slow first run alone is not a stable performance result.
A screenshot without identity, memory, and console context cannot diagnose a FLUX setup.
Memory tuning starts only after the base package is known and the initial attempt is recorded.
Increasing the value keeps more weights in dedicated VRAM but leaves less room for computation. Maximum is not the fastest setting.
Change in measured incrementsForcing another dtype can add conversion, double-quantization, or LoRA behavior you did not intend. Change it only as a named experiment.
Never stack guessesAsync and Shared can help some environments and hurt others. Record system RAM and shared-memory use before switching.
Compare same seed and workloadA base image loading does not prove Hires fix, LoRA, ControlNet, batch, or a larger aspect ratio fits.
Add one dependency per passBoth the current code and the official performance note warn that leaving almost no inference memory can force fallback, cause OOM, or make generation dramatically slower.
Inspect the official warning ↗Fix package identity before memory, and memory before optional workflow components.
Confirm the active models/Stable-diffusion path, finished download, supported extension, and top refresh control. A VAE or text encoder belongs to another inventory.
Treat this as a split-set failure. Verify CLIP-L—not CLIP-G or CLIP Vision—and the documented T5 file in models/text_encoder; select both in VAE / Text Encoder.
Final decoding implicates the VAE/package contract. Remove extra modules from a packaged checkpoint, or verify the exact FLUX ae.safetensors for a split setup. Disable extensions and retest.
Record whether the time is cold load or warm sampling. Restore Automatic/Queue/CPU, move GPU Weights below maximum, remove extra modules, and compare the exact same model, seed, dimensions, and Forge commit.
Keep batch 1, remove optional components, return to the documented native baseline, and lower GPU Weights in measured steps. If the smallest valid workload still exceeds written limits, reject that package.
Return to the current flux-preset contract: CFG 1 and the dev distribution’s Distilled CFG value. Describe the wanted result directly instead of importing an SDXL negative-prompt recipe.
Current Forge ignores Distilled CFG for recognized schnell models. Restore the schnell package’s documented low step count and guidance behavior.
Fix the seed and add one LoRA declared for that exact FLUX variant. Compare with and without it; record trigger, strength, console patches, low-bit mode, load time, VRAM, and RAM.
Troubleshoot a FLUX LoRA step by step →Read the dated official LoRA note ↗The answers separate variant, format, package, memory, and license decisions.
Yes. The inspected original Forge code contains a FLUX diffusion engine, FLUX and FLUX Schnell model detection, a flux UI preset, separate-module loading, and low-bit controls. Support for one architecture does not guarantee every conversion or add-on.
You need a current original Forge installation, one exact FLUX variant, its complete packaged or split file set, accepted model terms, enough local storage and memory for a measured baseline, and the matching flux interface controls.
Choose by workflow and terms. Dev uses distilled guidance and generally a longer sampling contract; schnell is a separate distilled variant whose publisher reference uses 1–4 steps. Their current publisher licenses also differ.
Current Forge recognizes both through its FLUX engine, but applies Distilled CFG to dev and ignores it for schnell. Use each exact distribution’s step, guidance, file, and license contract.
Not by itself. Dev or schnell is the model variant; NF4, FP8, GGUF, and full precision describe storage or computation choices and packaging. A fine-tune can add another identity layer.
There is no universal winner. Format, GPU generation, CUDA/PyTorch path, offload, encoders, VAE, resolution, batch, LoRA, and system RAM change the result. Shortlist a supported distribution and measure it locally.
Dated demonstrations report some low-memory configurations, but they are not a support guarantee. Test the smallest valid package at batch size 1 and record load success, dedicated VRAM, shared memory, RAM, time, and console output.
Not when the exact package documents those components as bundled. Do not attach separate modules just because another raw/GGUF tutorial uses them. Follow the package page.
Do not assume so. Official Forge’s split-layout guidance treats GGUF diffusion weights as one role and separately loads VAE, CLIP-L, and T5. Verify the exact distribution.
Primary packaged checkpoints and diffusion-model/GGUF files use models/Stable-diffusion by default; separate VAE uses models/VAE; CLIP-L and T5 use models/text_encoder. Path arguments can override these roots.
Use flux. It exposes VAE / Text Encoder, Diffusion in Low Bits, Swap Method, Swap Location, GPU Weights, and Distilled CFG controls with FLUX-oriented defaults. It does not convert or complete the selected model.
The current flux preset starts CFG at 1 and Distilled CFG at 3.5. Treat those as Forge interface defaults, then follow the exact dev distribution’s tested recipe and keep the values in saved infotext.
Current Forge’s FLUX engine explicitly ignores Distilled CFG for models recognized as schnell. Use the schnell distribution’s own guidance and step contract.
With the common dev setup at CFG 1, do not expect the classic SD negative-prompt workflow. Describe the wanted scene directly and follow the exact variant documentation rather than copying an SDXL recipe.
A split layout is missing the expected CLIP-L component or has the wrong encoder selected. CLIP-G and a CLIP vision encoder are not substitutes. Rebuild the documented set and select CLIP-L plus T5 in VAE / Text Encoder.
The preview can exist before VAE decoding. Return to the exact package contract: remove unnecessary modules for a packaged checkpoint, or verify the documented ae.safetensors for a split layout. Retest without extensions or LoRAs.
First record the exact model file, format, Forge commit, settings, and console. Do not set GPU Weights to the maximum: weights need VRAM, but computation also needs headroom. Compare cold load separately from warm generation.
No. Current Forge warns that using essentially all VRAM for weights leaves no compute headroom and can cause fallback, OOM, or severe slowdown. Start from the preset value and tune from measurements.
Do not. Prove the base package first, fix the seed, then add one LoRA declared compatible with the exact variant. Low-bit loading and LoRA precision can change patch time and memory behavior.
A family name is not legal clearance. The current BFL pages list FLUX.1-dev under the dev non-commercial license and FLUX.1-schnell under Apache-2.0; fine-tunes and conversions may add terms. Review the exact current licenses and obtain legal advice when needed.
Dated official experiments and community demonstrations locate risks, not permanent hardware promises.
FLUX/Schnell recognition, encoder roles, Distilled CFG behavior, preset defaults, file scanning, and memory-control warnings at dfdcbab.
Variant identity, reference usage, access conditions, displayed license, and limitations for BFL dev and schnell.
Packaged NF4/FP8, split raw/GGUF, LoRA, and performance posts document then-current experiments and layouts.
Five catalogued video transcripts and issue reports map setup confusion; their GPU thresholds, timing, and quality opinions are not universalized.