Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
JavaScript ai-agentsdashboarddeveloper-toolsdigital-twinquantified-selfsecond-brainself-hostedtailscale
/**
* Local-backend sampler knobs — frames, chunks (plus optional per-chunk
* prompt beats), fps, seed, steps, CFG scale, image strength, tiling and
* the audio flags. Always visible in the main form card (no disclosure)
* so the right-hand column can host Prompt from media.
*
* Presentational: all state and handlers are owned by the VideoGen page.
*/
import { Dice5 } from 'lucide-react';
import { FormField } from '../ui/FormField';
import {
frameOptionsForModel, fpsOptionsForModel, CHUNK_OPTIONS,
isModelAllowedForMode, supportsVideoAudioControls, supportsVideoAudioPromptControls,
CONTEXT_FRAME_OPTIONS, supportsContextWindow,
DEFAULT_SPEED_PROFILE_ID, speedProfilesForMode, selectedSpeedProfile,
videoChainChunkModes,
DEFAULT_DRAFT_DECODE_ID, draftDecodeOptionsForModel,
} from '../../lib/videoGenParams.js';
import { VIDEO_TILING_OPTIONS } from '../../lib/videoTilingOptions';
import { isLtx2FamilyRuntime } from '../../lib/runnerFamilies';
import { DEFAULT_VIDEO_STREAMING_MODE, VIDEO_STREAMING_MODE_OPTIONS } from '../../lib/videoStreamingMode.js';
import {
I2V_REFERENCE_MODE_OPTIONS, DEFAULT_I2V_REFERENCE_MODE,
normalizeI2vReferenceMode, runtimeSupportsI2vReferenceMode,
} from '../../lib/videoReferenceModes';
const inputCls = 'w-full bg-port-bg border border-port-border rounded-lg px-2 py-2 text-sm text-white focus:outline-none focus:border-port-accent disabled:opacity-50';
export default function AdvancedParamsPanel({
mode,
currentModel,
numFrames, onNumFramesChange,
chunks, onChunksChange, keyframesActive,
chunkPrompts = [], onChunkPromptChange, chainingActive = false,
contextFrames, onContextFramesChange,
fps, onFpsChange,
seed, onSeedChange, onRandomSeed,
batchSize = 1, onBatchSizeChange,
steps, onStepsChange,
guidanceScale, onGuidanceScaleChange,
speedProfileId = DEFAULT_SPEED_PROFILE_ID, onSpeedProfileChange,
draftDecode = DEFAULT_DRAFT_DECODE_ID, onDraftDecodeChange, draftDecodeLocked = false,
streamingMode = DEFAULT_VIDEO_STREAMING_MODE, onStreamingModeChange,
showStreamingMode = isLtx2FamilyRuntime(currentModel?.runtime),
imageStrength, onImageStrengthChange,
i2vReferenceMode = DEFAULT_I2V_REFERENCE_MODE, onI2vReferenceModeChange,
effectiveImageStrength = null,
tiling, onTilingChange,
disableAudio, onDisableAudioChange,
noMusic, onNoMusicChange,
idPrefix = '',
}) {
const fieldId = (id) => idPrefix ? `${idPrefix}-${id}` : id;
// a2v derives its length + audio track from the uploaded audio, so chunking
// and the audio flags don't apply there.
const showAudioFlags = mode !== 'a2v';
const showDisableAudio = showAudioFlags && supportsVideoAudioControls(currentModel);
const showPromptAudioControls = showAudioFlags && noMusic != null && supportsVideoAudioPromptControls(currentModel);
const audioDisabled = showDisableAudio && disableAudio;
const showFrames = !(mode === 'a2v' && currentModel?.audioDurationDriven === true);
// Chunk chaining seeds chunk N+1 from chunk N's last frame, so it needs i2v —
// the same predicate the picker uses, not a second reading of supportedModes.
const showChunks = mode !== 'a2v' && isModelAllowedForMode(currentModel, 'image');
// ltx2 extend conditions on the source's latent rather than a single frame,
// so image strength is meaningless for it.
const showImageStrength = mode === 'image' || (mode === 'extend' && !isLtx2FamilyRuntime(currentModel?.runtime));
// The reference-mode promise (#4874) is an image-mode question only, and only
// a runtime that carries per-image conditioning strength can keep anything but
// the default. Rendering it read-only elsewhere would imply a choice that does
// not exist, so the picker appears in image mode and explains itself when the
// selected model cannot honor the loose option.
const showReferenceMode = mode === 'image' && typeof onI2vReferenceModeChange === 'function';
const referenceMode = normalizeI2vReferenceMode(i2vReferenceMode);
// Which promises this model can actually keep. An unresolved `currentModel`
// means "the catalog has not loaded yet", NOT "anchor only" — the same
// deferral the form's snap-back makes, so the picker can't contradict the
// state it is rendering: collapsing to Anchor here would disable the control
// and claim "this model can only anchor" about a model nobody has seen.
const supportedReferenceOptions = currentModel
? I2V_REFERENCE_MODE_OPTIONS.filter((o) => runtimeSupportsI2vReferenceMode(currentModel.runtime, o.value))
: I2V_REFERENCE_MODE_OPTIONS;
// A controlled <select> must never hold a value with no matching <option>.
// A restored pick survives one render past a model switch — until the
// snap-back effect lands — so keep it listed for exactly that window.
const referenceOptions = supportedReferenceOptions.some((o) => o.value === referenceMode)
? supportedReferenceOptions
: [...supportedReferenceOptions, I2V_REFERENCE_MODE_OPTIONS.find((o) => o.value === referenceMode)].filter(Boolean);
const referencePromise = I2V_REFERENCE_MODE_OPTIONS.find((o) => o.value === referenceMode);
const referenceModeLocked = supportedReferenceOptions.length < 2;
// `0` is a legal strength (ignore the source entirely) and the retry editor
// hands it over as a NUMBER, so presence has to be tested explicitly — a
// `imageStrength || …` fallback renders the slider at 1 while the form still
// submits 0, which is the control lying about the render it will produce.
const hasExplicitStrength = imageStrength != null && imageStrength !== '';
const displayedImageStrength = hasExplicitStrength
? imageStrength
: (effectiveImageStrength != null ? effectiveImageStrength : 1);
const frameOptions = frameOptionsForModel(currentModel, numFrames);
const fpsOptions = fpsOptionsForModel(currentModel);
const samplerLocked = currentModel?.samplerLocked === true;
// Speed profiles (#4875). Only the profiles VALIDATED for what this request
// will actually run are offered: a profile the server would decline is a dead
// affordance that promises a speed-up the render then doesn't take. With none
// applicable the picker is hidden entirely rather than shown holding only
// "Quality".
//
// A CHAINED render is one clip whose chunks run in DIFFERENT modes, and the
// server applies a profile to it only when every chunk accepts one — so the
// gate takes the whole chunk-mode list. Without that, picking Fast with
// Chunks = 4 would grey out Steps/CFG and then quietly run the entire chain
// at the model default.
const speedProfileModes = videoChainChunkModes({
model: currentModel, mode, chaining: chainingActive, contextFrames,
});
// Preview-fidelity decode (#5423). The option list is server-declared and
// rides on the model entry, so a model with no draft decoder yields [] and
// renders NO control rather than a select with one real choice.
// A model the finish graph names as a DELIVERY target always decodes on its
// own decoder — `draftDecodeDeclineReason` refuses a draft request there
// before it even asks which decoder was meant. The caller reads that off the
// model list (`isDeliveryVideoModel`) and passes it down, so the control shows
// Full and says why rather than offering a choice the server would decline.
const draftDecodeOptions = draftDecodeOptionsForModel(currentModel);
const effectiveDraftDecode = draftDecodeLocked ? DEFAULT_DRAFT_DECODE_ID : draftDecode;
const activeDraftDecode = draftDecodeOptions.find((o) => o.id === effectiveDraftDecode) || null;
const speedProfiles = speedProfilesForMode(currentModel, speedProfileModes);
const showSpeedProfiles = !samplerLocked && speedProfiles.length > 0;
// Resolved against the SAME set the picker offers, so a model with two
// profiles can't lock Steps/CFG to one this request declines.
const activeSpeedProfile = showSpeedProfiles
? selectedSpeedProfile(speedProfileId, currentModel, speedProfileModes)
: null;
// A profile drives steps AND CFG together, exactly as a samplerLocked model
// does — leaving either editable would let the user half-override a
// validated schedule and get neither its speed nor its quality.
const samplerDriven = samplerLocked || !!activeSpeedProfile;
const samplerDrivenTitle = samplerLocked
? 'This validated Lightning profile locks its sampler settings.'
: activeSpeedProfile
? `The ${activeSpeedProfile.name} speed profile sets its own validated schedule. Switch to Quality to edit these.`
: undefined;
return (
<div className="border-t border-port-border pt-3 space-y-3">
<h2 className="text-xs font-medium text-gray-400 uppercase tracking-wide">Advanced</h2>
<div className="grid grid-cols-2 sm:grid-cols-3 gap-3">
{showFrames && <FormField label="Frames" labelClassName="block text-xs font-medium text-gray-400 mb-1">
<select
value={numFrames}
onChange={(e) => onNumFramesChange(Number(e.target.value))}
className={inputCls}
>
{frameOptions.map((f) => (
<option key={f} value={f}>
{f} ({(f / fps).toFixed(1)}s @ {fps}fps){f === currentModel?.defaultFrames ? ' · default' : ''}
</option>
))}
</select>
{numFrames > 241 && isModelAllowedForMode(currentModel, 'extend') && (
<p className="text-[10px] text-gray-500 leading-snug mt-1">
Past 241 frames a single-pass render may swap or OOM at 48 GB. For reliable longer clips, render up to ~10s and then use <strong>Extend</strong> on the result — it conditions on the source's full latent rather than a single last frame.
</p>
)}
</FormField>}
{showChunks && (
<div>
<label htmlFor={fieldId('chunks-select')} className="block text-xs font-medium text-gray-400 mb-1" title="Chain N renders end-to-end. Each chunk's last frame seeds the next, then they're stitched into one clip. Wall time scales linearly with chunks.">
Chunks
</label>
<select
id={fieldId('chunks-select')}
value={keyframesActive ? 1 : chunks}
onChange={(e) => onChunksChange(Number(e.target.value))}
disabled={keyframesActive}
title={keyframesActive ? 'Multi-keyframe renders anchor a single clip — chunking is unavailable.' : undefined}
className={`${inputCls} disabled:cursor-not-allowed`}
>
{CHUNK_OPTIONS.map((n) => (
<option key={n} value={n}>
{n === 1 ? '1 (single)' : `${n} (~${((n * numFrames) / fps).toFixed(0)}s total)`}
</option>
))}
</select>
</div>
)}
{/* Continuation context window. Only meaningful once the request
really chains AND the runtime has an extend pipeline to feed the
window to — everywhere else the server seeds the next chunk from
a single last frame regardless, so showing the control would be
offering a knob that does nothing. */}
{chainingActive && supportsContextWindow(currentModel) && (
<div>
<label htmlFor={fieldId('context-frames-select')} className="block text-xs font-medium text-gray-400 mb-1" title="How much of the previous chunk each new chunk sees. A window carries the scene's motion across the seam; a single last frame gives the model a pose with no velocity, so movement stalls and restarts at every join.">
Continuity
</label>
<select
id={fieldId('context-frames-select')}
value={contextFrames}
onChange={(e) => onContextFramesChange(Number(e.target.value))}
className={inputCls}
>
{CONTEXT_FRAME_OPTIONS.map((n) => (
<option key={n} value={n}>
{n === 0 ? 'Last frame only' : `${n} frames (~${(n / fps).toFixed(1)}s)`}
</option>
))}
</select>
<p className="text-[10px] text-gray-500 leading-snug mt-1">
Bigger windows hold motion better across joins but add render time per chunk.
</p>
</div>
)}
{/* Per-chunk prompt beats (#3695). Only shown when the request really
chains — the parent derives `chainingActive` from the same
predicate that decides whether `chunks` is submitted at all, so
this editor can never appear for a mode the server pins to one
chunk. One row per LIVE chunk; the parent keeps text for chunks
beyond the current count, so lowering then raising the count
restores what was typed. */}
{chainingActive && (
<div className="col-span-2 sm:col-span-3">
<p className="block text-xs font-medium text-gray-400 mb-1">
Per-chunk beats <span className="font-normal text-gray-500">(optional — blank uses the main prompt)</span>
</p>
<div className="space-y-1.5">
{Array.from({ length: chunks }, (_, i) => (
<div key={i} className="flex items-center gap-2">
<label htmlFor={fieldId(`chunk-prompt-${i}`)} className="text-[11px] text-gray-500 w-12 shrink-0">
Chunk {i + 1}
</label>
<input
id={fieldId(`chunk-prompt-${i}`)}
type="text"
value={chunkPrompts[i] || ''}
onChange={(e) => onChunkPromptChange(i, e.target.value)}
placeholder="Same as main prompt"
className={inputCls}
/>
</div>
))}
</div>
<p className="text-[10px] text-gray-500 leading-snug mt-1">
Each chunk still continues from the previous chunk's last frame — beats steer what happens next rather than starting a new shot.
</p>
</div>
)}
<FormField label="FPS" labelClassName="block text-xs font-medium text-gray-400 mb-1">
<select
value={fps}
onChange={(e) => onFpsChange(Number(e.target.value))}
className={inputCls}
>
{fpsOptions.map((f) => <option key={f} value={f}>{f}</option>)}
</select>
</FormField>
{onBatchSizeChange && (
<div>
<label htmlFor={fieldId('video-batch-size')} className="block text-xs font-medium text-gray-400 mb-1">Renders in batch</label>
<select id={fieldId('video-batch-size')} className={inputCls}
value={currentModel?.supportsWarmBatch && !chainingActive ? batchSize : 1}
disabled={!currentModel?.supportsWarmBatch || chainingActive}
onChange={(event) => onBatchSizeChange(Number(event.target.value))}>
{Array.from({ length: 20 }, (_, index) => <option key={index + 1} value={index + 1}>{index + 1}</option>)}
</select>
<p className="mt-1 text-xs text-gray-500">
{!currentModel?.supportsWarmBatch ? 'Warm batching is available for local MiniMax H3 MLX models.'
: chainingActive ? 'Use one chunk to batch separate videos.'
: 'Sequential videos reuse the model and prompt. Blank seed: random each time. Number: increment per render.'}
</p>
</div>
)}
<div>
<label htmlFor={fieldId('video-seed')} className="block text-xs font-medium text-gray-400 mb-1">Seed</label>
<div className="flex items-center gap-1">
<input
id={fieldId('video-seed')}
type="number"
value={seed}
onChange={(e) => onSeedChange(e.target.value)}
placeholder="Random"
className={`flex-1 ${inputCls}`}
/>
<button
type="button"
onClick={onRandomSeed}
className="p-2 text-gray-400 hover:text-white border border-port-border rounded-lg hover:bg-port-border/50 disabled:opacity-50 min-h-[40px] min-w-[40px] flex items-center justify-center"
title="Use a random seed for each render" aria-label="Randomize seed"
>
<Dice5 className="w-4 h-4" />
</button>
</div>
</div>
{draftDecodeOptions.length > 0 && (
<FormField label="Decode" labelClassName="block text-xs font-medium text-gray-400 mb-1">
<select
value={activeDraftDecode ? effectiveDraftDecode : DEFAULT_DRAFT_DECODE_ID}
disabled={draftDecodeLocked}
onChange={(e) => onDraftDecodeChange?.(e.target.value)}
className={inputCls}
>
{draftDecodeOptions.map((option) => (
<option key={option.id} value={option.id}>
{option.label}{option.sizeLabel ? ` · ${option.sizeLabel}` : ''}
</option>
))}
</select>
<p className="mt-1 text-xs text-gray-500">
{draftDecodeLocked
? `${currentModel?.name || 'This model'} is a delivery model — final and Finish renders always decode on the full decoder.`
: (
<>
{activeDraftDecode?.description || ''}
{activeDraftDecode && activeDraftDecode.id !== DEFAULT_DRAFT_DECODE_ID
? ' Finish and delivery renders always use the full decoder.'
: ''}
</>
)}
</p>
</FormField>
)}
{/* Block streaming (#6499) — LTX-2/2.5 MLX only, every other runtime
ignores the flag by construction. Auto initially streams on a
tight-memory machine; the render bridge inspects the pinned
pipeline and this machine's physical RAM, so the client offers no
per-model capability table for it. */}
{showStreamingMode && (
<FormField label="Memory" labelClassName="block text-xs font-medium text-gray-400 mb-1">
<select
value={streamingMode}
onChange={(e) => onStreamingModeChange?.(e.target.value)}
className={inputCls}
>
{VIDEO_STREAMING_MODE_OPTIONS.map((option) => (
<option key={option.value} value={option.value}>{option.label}</option>
))}
</select>
<p className="mt-1 text-xs text-gray-500">
{VIDEO_STREAMING_MODE_OPTIONS.find((o) => o.value === streamingMode)?.description || ''}
</p>
</FormField>
)}
{showSpeedProfiles && (
<FormField label="Speed" labelClassName="block text-xs font-medium text-gray-400 mb-1">
<select
value={activeSpeedProfile ? speedProfileId : DEFAULT_SPEED_PROFILE_ID}
onChange={(e) => onSpeedProfileChange(e.target.value)}
className={inputCls}
>
<option value={DEFAULT_SPEED_PROFILE_ID}>Quality · default</option>
{speedProfiles.map((p) => (
<option key={p.id} value={p.id}>
{p.name}{p.speedupLabel ? ` · ${p.speedupLabel}` : ''}
</option>
))}
</select>
</FormField>
)}
<FormField
label={<>Steps {currentModel?.steps && `(default: ${currentModel.steps})`}</>}
labelClassName="block text-xs font-medium text-gray-400 mb-1"
>
<input
type="number" min={1} max={150}
value={activeSpeedProfile ? '' : steps}
onChange={(e) => onStepsChange(e.target.value)}
disabled={samplerDriven}
title={samplerDrivenTitle}
placeholder={String(activeSpeedProfile?.steps ?? currentModel?.steps ?? 25)}
className={inputCls}
/>
</FormField>
<FormField
label={<>CFG Scale {currentModel?.guidance != null && `(default: ${currentModel.guidance})`}</>}
labelClassName="block text-xs font-medium text-gray-400 mb-1"
>
<input
type="number" min={0} max={20} step={0.5}
value={activeSpeedProfile ? '' : guidanceScale}
onChange={(e) => onGuidanceScaleChange(e.target.value)}
disabled={samplerDriven}
title={samplerDrivenTitle}
placeholder={String(activeSpeedProfile?.guidance ?? currentModel?.guidance ?? 3.0)}
className={inputCls}
/>
</FormField>
{samplerLocked && (
<p className="col-span-2 sm:col-span-3 text-[10px] text-port-accent leading-snug">
{currentModel?.samplerNote
|| `Lightning keeps its validated ${currentModel.steps}-step sampler, CFG ${currentModel.guidance}, and model-specific schedule locked.`}
</p>
)}
{showReferenceMode && (
<div className="col-span-2 sm:col-span-3">
<label htmlFor={fieldId('video-reference-mode')} className="block text-xs font-medium text-gray-400 mb-1">Reference mode</label>
<select
id={fieldId('video-reference-mode')}
value={referenceMode}
disabled={referenceModeLocked}
onChange={(e) => onI2vReferenceModeChange(e.target.value)}
className={inputCls}
>
{referenceOptions.map((o) => <option key={o.value} value={o.value}>{o.label}</option>)}
</select>
<p className="text-[10px] text-gray-500 leading-snug mt-1">{referencePromise?.promise}</p>
{referenceModeLocked && (
<p className="text-[10px] text-gray-500 leading-snug mt-1">
{currentModel?.name || 'This model'} can only anchor a reference. Pick an LTX-2.5 model to loosen it.
</p>
)}
</div>
)}
{activeSpeedProfile && (
<p className="col-span-2 sm:col-span-3 text-[10px] text-port-accent leading-snug">
{activeSpeedProfile.description}
{' '}Steps and CFG follow the profile ({activeSpeedProfile.steps}
{activeSpeedProfile.stage2Steps != null ? `+${activeSpeedProfile.stage2Steps}` : ''} steps, CFG {activeSpeedProfile.guidance}).
{' '}If a lever it needs is missing on this machine, the render still runs and says so rather than claiming the speed-up.
</p>
)}
{showImageStrength && (
<div className="col-span-2 sm:col-span-3">
<div className="flex items-center justify-between gap-3 mb-1">
<label htmlFor={fieldId('video-image-strength')} className="block text-xs font-medium text-gray-400">Image Strength</label>
<span className="text-[11px] text-gray-500">{hasExplicitStrength || effectiveImageStrength != null ? String(displayedImageStrength) : '1.0'}</span>
</div>
<input
id={fieldId('video-image-strength')}
type="range" min={0} max={1} step={0.05}
value={displayedImageStrength}
onChange={(e) => onImageStrengthChange(e.target.value)}
className="w-full accent-port-accent"
title="Higher values preserve the source frame more strongly"
/>
</div>
)}
{currentModel?.supportsTiling !== false && (
<FormField className="col-span-2 sm:col-span-3" label="Tiling" labelClassName="block text-xs font-medium text-gray-400 mb-1">
<select
value={tiling}
onChange={(e) => onTilingChange(e.target.value)}
className={inputCls}
>
{VIDEO_TILING_OPTIONS.map((t) => <option key={t.value} value={t.value}>{t.label}</option>)}
</select>
</FormField>
)}
{showDisableAudio && (
<label className="col-span-2 sm:col-span-3 flex items-center gap-2 text-xs text-gray-400 cursor-pointer">
<input
type="checkbox"
checked={disableAudio}
onChange={(e) => onDisableAudioChange(e.target.checked)}
className="rounded"
/>
Disable audio (LTX-2 only — speeds up generation)
</label>
)}
{showPromptAudioControls && (
<label
className={`col-span-2 sm:col-span-3 flex items-center gap-2 text-xs cursor-pointer ${audioDisabled ? 'text-gray-600 cursor-not-allowed' : 'text-gray-400'}`}
title="The model conditions generated audio on the prompt — appending 'no music, no soundtrack' at submit time pushes it toward ambient/diegetic sound only"
>
<input
type="checkbox"
checked={noMusic}
disabled={audioDisabled}
onChange={(e) => onNoMusicChange(e.target.checked)}
className="rounded"
/>
No music — keep ambient/diegetic sound only
</label>
)}
</div>
</div>
);
}