Parameter reference

Every flag, findable in two clicks

Search by name, or filter by what you're tuning. Defaults that shift between releases are flagged version-dependent - always confirm against your ff-run --help.

43 flags

--execution-providers

list= cuda

Which executors run the models (cuda, cpu, directml, openvino, rocm, coreml, tensorrt).

--execution-thread-count

int= version-dependent

Worker threads per request; keep near your core count.

--execution-queue-count

int= low

Bounded queued frames between stages; higher uses more RAM.

--vram-limitvary

string= version-dependent

Cap per-provider GPU memory so models share one card.

--system-memory-limitvary

string= version-dependent

Cap host RAM used by the pipeline.

--face-detector-size

choice= ~512x512

Detection resolution; larger = more recall, more cost.

--face-detector-score

float 0-1= ~0.5

Min confidence for a detection; lower = more faces.

--face-detector-angles

list= 0,90,180,270

Internal rotations to catch tilted faces; more = slower.

--face-detector-model

choice= version-dependent

yoloface | scrfd | retinaface (plus 'many' in newer builds).

--face-landmarker-model

choice= 2dfan4

Landmark network for tracking/alignment (2dfan4, pipnet in newer).

--face-landmarker-score

float 0-1= ~0.5

Min confidence for landmarks; lower = more permissive.

--face-recognizer-model

choice= arcface_w600k_r50

Embedding network used for matching/verifying faces.

--face-recognizer-distance

choice= cosine_distance

Distance metric for comparing face embeddings.

--face-recognizer-metric

choice= max

How to aggregate multi-face comparisons (max/min/average).

--face-recognizer-thresholdvary

float= ~0.4 cosine

Cutoff for a 'match'; lower = stricter.

--face-swapper-model

choice= inswapper_128

inswapper_128 | simswap_256/512 | ghost_256 | uniface_256.

--face-swapper-pixel-boostvary

choice= off

Second-pass upscale of the swap region for sharper output.

--face-enhancer-model

choice= codeformer

codeformer | gfpgan_1.x | restoreformer(+_plus_plus) | gpen_bfr_*.

--face-enhancer-blend

float 0-1= ~0.8

Strength of the enhanced face over the original.

--face-enhancer-weight

float 0-1= ~0.5

Fidelity vs. restoration for codeformer/gfpgan.

--frame-enhancer-model

choice= real_esrgan_x4plus

Whole-frame upscaler (real_esrgan_*, span_kendata_x4).

--frame-enhancer-blend

float 0-1= ~0.8

Strength of the upscaled frame over the original.

--lip-syncer-model

choice= wav2lip

wav2lip | wav2lip_gan - syncs lips to --audio.

--lip-syncer-blend

float 0-1= ~0.8

Strength of the synced mouth over the original.

--audio

path= -

Audio track that lip_syncer animates the mouth to.

--expression-restorer-modelvary

choice= live_portrait

Drives expression transfer.

--expression-restorer-factorvary

float 0-1= version-dependent

How strongly the driving expression is applied.

--face-debugger-items

list= face

Overlays: face (bbox), face_landmark, face_mask, face_landmark_mask.

--face-mask-types

list= occlusion

box | occlusion | region - how the swap region is computed.

--face-mask-blur

float 0-100= version-dependent

Feathers the mask edge for blending.

--face-mask-padding

t r b l px= 0 0 0 0

Expand (+) or contract (-) the mask.

--face-mask-regions

list= skin, ...

Per-feature regions when using region mask (eyes, brows, lips, nose, hair, neck, ...).

--output-video-codec

choice= libx264

hw: h264_nvenc, hevc_nvenc, h264_amf, hevc_amf, h264_qsv, hevc_qsv. sw: libx264, libx265, libsvtav1, libvpx-vp9.

--output-video-quality

int 0-51= 18

CRF; lower = higher quality/bigger file. Guide sweet spot 18 - 20.

--output-video-preset

choice= veryfast

Speed/compression trade for the encoder.

--output-video-bitrate

int= off

Absolute bitrate cap when targeting a file size.

--output-video-grain

float= off

Adds film grain to mask banding and match source.

--temp-frame-format

choice= png

Intermediate format (png | jpg). png = lossless, heavier disk.

--keep-temp

flag= off

Keep temp frames on disk for inspection.

--trim-frame-start

int= 0

First frame of the input to process.

--trim-frame-end

int= last

Last frame of the input to process.

--temp-frame-quality

int= version-dependent

Compression for lossy temp frames (jpg).

--output-video-fps

float= -

Output frame rate; derive from input if unset.

Categories mirror the concerns you actually tune: Performance, Quality, Tracking, Face, Masking, Encoding, Audio. VRAM figures and exact defaults stay qualitative here on purpose.