A physical gate sets adaptation per token and head.
Constant gains instead of Ψ cost 0.81 dB and 6.09 ΔEITP.
LumaFlux (arXiv:2604.02787)
Scene context from SigLIP, and a monotone tone decoder
PCM sets one scale and shift per frame from pooled SigLIP tokens.
Ŷ = RQS(Yout; ξ, η, s)(4)
x̂hdr = MRGB([Ŷ, Û, V̂]) in PQ BT.2020
25 spline parameters per frame, all positive.
So decoded luma keeps its ordering.
Removing it costs 9.69 dB.
8 bins: 8 widths ξ, 8 heights η, 9 knot slopes s, from a 21.9K-parameter head that starts at identity.
LumaFlux (arXiv:2604.02787)
A 314,396-pair corpus of user and professional video
314,396SDR-HDR pairs in PQ BT.2020
2,092source videos
144GPU-hours on four GH200 GPUs, about 36 hours
Source
What it adds
CHUG (Saini et al., ICIP 2025)
856 user-generated references, 846 train and 10 reserved
LIVE-TMHDR
40 references with 120 expert-graded SDR clips
HDRTV1K
1,235 native training pairs
Sol Levante
one HDR10 master
Synthetic SDR
307,752 rows: eight tone-mapping operators, three CRF settings
Mix
171,384 user-generated (UGC), 143,012 professional (PGC)
All HDR is 16-bit PQ BT.2020. A manifest and code regenerate the corpus in about four CPU hours.
Training
Setting
Loss
flow matching, two L1 terms in linear light, spline smoothness
Crops
5122, 1,024 tokens
Schedule
100,000 steps, global batch 16, learning rate 2e-4
Trained
71.3M parameters; the backbone stays frozen
In-domain checkpoint
HDRTV1K only, 50,000 steps, adapters only
The in-domain checkpoint is labeled wherever it appears. Every other result uses the main, mixed-corpus model.
LumaFlux (arXiv:2604.02787)
LumaFlux is best in every column of the primary track
LumaEval synthetic track, 100 held-out frames under eight tone-mapping degradations, one scoring stack
Method
PU21-PSNR ↑
PU21-SSIM ↑
ΔEITP ↓
HDR-LPIPS ↓
HDR-VDP-3 ↑
FR-HIDRO ↓
BT.2446c
23.74
0.8291
61.60
0.307
5.611
0.753
Reinhard operator
23.62
0.8286
61.76
0.295
5.726
0.753
HDRTVNet++
23.03
0.8192
66.67
0.328
5.417
0.703
FMNet
22.86
0.8176
67.92
0.330
5.412
0.724
ITM-LUT
22.92
0.8142
67.41
0.315
5.496
0.761
VAE + RQS (control)
18.54
0.7647
111.67
0.381
3.673
0.924
LumaFlux (ours)
24.23
0.8294
56.54
0.269
5.812
0.631
Macro-averages over the same 100 IDs and 78 video-variant groups. Bold is best, underline second best. PU21-PSNR-Y is in the full table (backup).
24.23 dBPU21-PSNR, 0.49 dB above the best baseline
56.54ΔEITP, 5.06 lower
0.631FR-HIDROVQA, from 0.703 for HDRTVNet++
FR-HIDROVQA is a full-reference distance on HIDRO-VQA features (Saini et al., WACV 2024).
LumaFlux (arXiv:2604.02787)
Three scenes, every method
SDR input with the patch box, then each method's crop. Full-frame PU21-PSNR is above each crop, best in bold.
On Sunset LumaFlux reaches 28.22 dB; the next best, BT.2446c, reaches 24.67.
LumaFlux (arXiv:2604.02787)
In the ablation, the tone decoder matters most
Row
Configuration
PSNR (Δ vs g)
ΔEITP (Δ vs g)
(c)
PGA, no spectral gate
28.57 (−0.62)
33.48 (+5.66)
(d)
PGA + spectral gate
29.56 (+0.37)
28.38 (+0.57)
(f)
full design, seed-0 run 1
28.67 (−0.52)
35.00 (+7.18)
(g)
full design, seed-0 run 2 (ours)
29.19 (ref.)
27.81 (ref.)
(h)
full, linear spline instead of RQS
28.74 (−0.45)
32.16 (+4.35)
(i)
full, no RQS decoder
19.50 (−9.69)
261.78 (+233.97)
(j)
full, no timestep-layer modulation
28.38 (−0.81)
33.90 (+6.09)
20k steps on a 55-video pool, the same 82 held-out frames. Deltas are relative to row (g). Rows (a), (b), (e) are in the full table (backup).
No RQS, row (i): PSNR falls 9.69 dB.
A linear spline, row (h), costs 0.45 dB.
The spectral gate, (c) to (d), adds 0.98 dB.
Rows (f) and (g) are one configuration, 0.52 dB apart, so smaller gaps are noise.
LumaFlux (arXiv:2604.02787)
Video without temporal training
Per-frame mean PU21 luma, first 120 frames of four clips, relative to each reference mean. Stabilized LumaFlux only.
Method
Bonfire
02c3
055e
0392
0237
Mean
Reference
0.000
0.000
0.000
0.000
0.000
0.000
BT.2446c
7.813
1.411
0.450
0.063
0.124
1.972
HDRTVNet++
3.005
0.207
0.147
0.442
1.257
1.012
FMNet
3.853
0.202
0.074
0.579
1.124
1.166
LumaFlux
1.992
0.691
2.500
0.968
0.569
1.344
LumaFlux + stabilization
2.109
0.469
1.676
0.879
0.362
1.099
|flickerpred − flickerref|, 106 PU21-energy units, lower is closer. 02c3 to 0237 are CHUG clips (Saini et al., ICIP 2025).
Draw the bridge noise once and reuse it for every frame.
Smooth the spline with an EMA, θ̄f = 0.8 θ̄f−1 + 0.2 θf.
Signed excess flicker falls 53.3%, from 547.4 to 255.5 (×103).
Raw flicker energy stays within 3.27% of the reference.
LumaFlux (arXiv:2604.02787)
What holds up, and what falls short
Holds up under compression
PSNR change from QP 27 on the 117 HDRTV1K pairs, HEVC intra coding of the SDR input. Flatter is better.
From QP 27 to 42 the main model loses 0.55 dB; HDRTVNet++ loses 2.77 and FMNet 2.71.
Falls short
LumaFlux
Reference point
HDRTV1K, in-domain checkpoint
33.34 dB
38.60 dB, best published
HDRTV1K, main model zero-shot
27.52 dB
same protocol
Mean ITP chroma, five scenes
60.65
49.80, the references
Seconds per 1080p frame
5.282
0.367 to 1.708, single pass
The VAE is an SDR codec, so heavily clipped detail stays lost.
Colors run oversaturated, and in-domain CNNs stay ahead on HDRTV1K.
LumaFlux (arXiv:2604.02787)
Takeaways
Measurements can steer a frozen flow transformer. 0.57% of weights trained, no prompt.
A monotone tone field keeps luminance ordering. Worth 9.69 dB.
One checkpoint leads every primary metric and does video. 24.23 dB on LumaEval, flicker down 53.3%.
Saini et al., arXiv:2604.02787 · github.com/shreshthsaini/LumaFlux · shreshthsaini.github.io/LumaFlux
Next, Rectified-CFG++ (Saini et al., NeurIPS 2025) changes how guidance is applied.
LumaFlux (arXiv:2604.02787)
Part 3 of 3 · Generative guidance
Rectified-CFG++, geometry-aware flow guidance
How do we guide a rectified flow model without pushing it off its manifold?
Saini, Gupta and Bovik, NeurIPS 2025
The question
Why does guidance break on the best image generators?
Flux, SD3 and SD3.5 sample a rectified flow with a deterministic ODE. Classifier-free guidance (CFG) pushes their trajectories off the manifold, and nothing brings them back.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Why we need guidance at all
Without it, a flow model gives generic images that only loosely match the prompt
No guidance. Dark, muddy, weak on the prompt.
CFG. Follows the prompt, but oversaturates.
Ours. Follows the prompt and stays natural.
CFG was built for stochastic diffusion samplers (DDPM, SDXL). Flow models sample a deterministic ODE.
Flux-dev, same prompt and seed
No guidance
CFG, ω = 3.5
Ours, λ = 3.5
SD3.5, same prompt and seed
No guidance
CFG, ω = 3.5
Ours, λ = 3
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Fresh noise at every step pulls the sample back toward the data
Flow ODE Flux, SD3, SD3.5
dx = vθ(x, t) dt
Nothing pulls it back. Step t+1 starts off the manifold and the error compounds
So we need guidance designed for deterministic ODE samplers.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Why CFG fails and how we fix it, geometrically
Ideal flow. The velocity stays tangent to the manifold.
CFG. With ω > 1 the step lands off the manifold, for good.
Ours. Predict a midpoint with vc, evaluate guidance there, add a bounded correction.
Left to right: ideal flow, CFG, Rectified-CFG++. The step-by-step version is in the backup slides.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Extrapolation versus interpolation
This one difference explains why CFG fails and Rectified-CFG++ works
CFG extrapolates (ω > 1)
ω = 3 means three times the distance from vu to vc, into territory the model never learned.
Ours interpolates (0 ≤ α ≤ 1)
The base is always vc, on the manifold. The correction α(t) Δvmid is bounded and decays to 0 near the data.
CFG runs ω from 3 to 9. Ours keeps α between 0 and 1.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
The two algorithms side by side
Highlighted lines are what changes
Standard CFG
for t = T down to 1:
vc = vθ(xt, t, y)
vu = vθ(xt, t, ∅)
v̂ = vu + ω(vc − vu)extrapolation
xt−1 = xt + Δt v̂
2 network evaluations (NFE) per step. Nothing corrects the overshoot.
Rectified-CFG++
for t = T down to 1:
vc = vθ(xt, t, y)
x̃mid = xt + (Δt/2) vcpredictor
vcmid = vθ(x̃mid, t − Δt/2, y)
vumid = vθ(x̃mid, t − Δt/2, ∅)
v̂ = vc + α(t)(vcmid − vumid) interpolation
xt−1 = xt + Δt v̂
3 NFE per step. α(t) = λ(1−t)γ decays to 0 near the data.
The base is always vc, guidance comes from the midpoint, and α(t) fades out.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Theoretical guarantees
Four mild assumptions give three bounds. No earlier guidance method has any.
Result
Bound
In words
Lemma 1 guidance stability
‖Δvθt−Δt/2 − Δvθt(xt)‖ ≤ L Vmax Δt
Guidance at the midpoint differs from guidance at the current point by only O(Δt)
Proposition 1 bounded single-step perturbation
‖x̂t−1 − x̃t−1‖ ≤ α(t)B Δt
Deviation from the conditional path is set by α(t), so it vanishes near the data
Proposition 2 distributional deviation
DKL(p̂0 ‖ p0) ≤ Cs (B + L Vmax Δt) ∫α(τ) dτ
Total KL divergence is set by the integrated guidance ∫α(τ)dτ
Assumptions. (A1) vθ(x,t,y) and vθ(x,t,∅) are L-Lipschitz in x. (A2) Bounded guidance direction, ‖Δvtθ(x)‖ ≤ B. (A3) The schedule α(t) is bounded and integrable. (A4) Bounded conditional velocity, ‖vct‖ ≤ Vmax.
One knob, ∫α(τ)dτ, controls how far the output can drift.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
CFG versus ours, property by property
Lemma 2, manifold-faithful corrector
dist(x̂, Mt−Δt) ≤ C ε Δt
Distance to the manifold is bounded by the training error ε alone.
Integrated guidance
4 to 16× lower KL divergence
Property
CFG
Ours
Per-step perturbation
(ω−1) B Δt constant, never vanishes
α(t) B Δt goes to 0 near the data
Manifold distance
Unbounded errors accumulate
O(ε Δt) training error only
Integrated guidance
ω−1 ≈ 2 to 8
λ/(γ+1) ≈ 0.5
KL(p̂0 ‖ p0)
O(ω−1)
O(0.5)
Late-step behavior
Same ω everywhere, destroys fine detail
α(t) goes to 0, preserves fine detail
Guidance evaluation
At the current point (possibly off the manifold)
At the predicted midpoint (on the manifold)
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Watching the artifacts grow, or not
CFG ω = 2.5
Ours λ = 0.5
CFG, ω = 2.5
Ours, λ = 0.5
Each strip shows 7 denoising steps, from noise on the left to the final image on the right.
CFG drifts by step 2 or 3 and every later step amplifies it. Ours looks natural throughout.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Baseline, CFG and Rectified-CFG++ on three Flux prompts
Rows, top to bottom: Flux baseline with no guidance, CFG, Rectified-CFG++.
Cactus. CFG draws a cartoon with an emoji face.
Dog and moon gate. CFG loses the stone arch and paints mountains.
Parrot. CFG turns the bird into a psychedelic abstraction.
Ours keeps the realism and the structure in all three.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
CFG versus ours on Flux-dev, eight pairs
CFG
Ours
CFG
Ours
CFG
Ours
CFG
Ours
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Text rendering, the stop sign test
Text needs pixel-precise strokes in the last denoising steps
CFG
Rectified-CFG++
Prompt: "A stop sign with 'ALL WAY' written below it."
CFG. STOP warps and ALL WAY is unreadable.
Ours. α(t) is 0 in the last steps, so the strokes stay crisp.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Results on MS-COCO 10K
Model
Method
FID ↓
CLIP ↑
ImgRwd ↑
Pick ↑
HPS ↑
Lumina
CFG
26.93
0.261
0.547
21.03
0.253
Lumina
Ours
22.49
0.268
0.621
21.19
0.259
SD3
CFG
26.33
0.273
0.684
21.48
0.264
SD3
Ours
24.68
0.278
0.753
21.62
0.269
Flux
CFG
37.86
0.285
0.892
22.04
0.279
Flux
Ours
32.23
0.289
0.961
22.18
0.283
CFG versus Rectified-CFG++ on three flow models. Bold is best. ImgRwd is ImageReward, Pick is PickScore.
Guidance (SD3.5)
FID ↓
ImgRwd ↑
CLIP ↑
HPSv2 ↑
CFG
67.71
1.053
0.352
0.294
CFG-Zero*
68.39
0.995
0.346
0.288
APG
67.23
1.075
0.351
0.294
Rectified-CFG++
67.15
1.085
0.351
0.296
Compared with the other guidance methods on SD3.5.
Best or tied on every metric, for every model.
−5.6FID on Flux
−4.4FID on Lumina
4models tested
0extra training
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
SigLIP-2, for its strong semantic priors and multimodal compatibility
Contrastive, because HDR information is what HDR has and SDR lacks
10-bit PQ input, because tone mapping destroys the signal we need
Captions from Qwen2.5-VL-72B, written to describe quality
SigLIP-2 contrastive fine-tuning with matched HDR-SDR pairs and quality-aware captions
PhD defenseHDR-Q, Saini et al., CVPR 2026
Training pipeline and reward design
Stage 1, modality alignment
Full HAPO with γ = 0.5 on a curated Beyond8Bits subset of matched HDR-SDR pairs. The goal is to teach the model to attend to HDR-specific information. Both stages use RL; there is no SFT stage.
Stage 2, quality calibration
The full Beyond8Bits corpus, with the score reward as the primary signal and a reduced γ. The goal is prediction accuracy while keeping the HDR grounding from stage 1.
Why two stages
Modality alignment and quality calibration compete. Aggressive contrastive training degrades absolute accuracy, and pure quality training allows modality neglect. Doing them in sequence resolves the tension.
Composite reward
R = wfmt · Rfmt + wsc · Rscore + wself · Rself
Rfmt is binary, 1 for valid <think> and <answer> tags
Rscore is Gaussian, exp(−(ŝ−s*)2/2σ2) with σ = 3, smooth and differentiable
Rself is majority-vote consistency across K = 8 completions
Setup. Ovis2.5 (9B) with rank-4 LoRA, K = 8, ε = 0.1, 4 × H200, BF16, T = 8 frames, native 10-bit PQ input.
PhD defenseHDR-Q, Saini et al., CVPR 2026
HAPO vs GRPO during training
Token entropy during training. GRPO collapses; HAPO keeps healthy exploration.
GRPO failure mode
Entropy drops, generation turns deterministic, the model ignores the visual modality and takes text-context shortcuts.
HAPO stabilization
Dual entropy holds H ≈ 0.33 at quality-critical positions, so exploration survives.
Reasoning efficiency
Chain of thought shrinks from 168 to 137 tokens (−18%). Boilerplate goes, quality observations stay.
PhD defenseHDR-Q, Saini et al., CVPR 2026
What we learned, and what remains
What we learned
Modality neglect, rather than encoder capacity, is the bottleneck. The contrastive KL mechanism is necessary and sufficient for HDR grounding.
Token-level entropy is a diagnostic signal. Rising entropy at decision points means better visual grounding; collapsing entropy means text shortcuts.
Two-stage RL beats SFT followed by RL. Both stages use RL, so the HDR grounding is there from initialization.
Remaining challenges
Temporal reasoning is limited to T = 8 sampled frames, so long-range temporal quality patterns may be missed.
Extreme out-of-distribution content (synthetic HDR, gaming HDR) is thin in the training data.
Inference cost. An MLLM is slower than lightweight VQA models, so this is not real-time yet.
Luminance. PQ reaches 1,000 cd/m2; SDR gamma stops near 100. Everything above the ceiling becomes flat white.
Color. BT.2020 is squeezed into BT.709, then 10-bit PQ codes become 8-bit gamma codes and a codec quantizes them.
LumaFlux (arXiv:2604.02787)
Why the inverse is ill-posed
Clipping is many-to-one
Every luminance above the ceiling lands on the same code value, so very different highlights become one flat white.
Gamut mapping loses chroma
Distinct BT.2020 colors collapse onto the BT.709 boundary. The SDR code does not say how far outside they were.
Quantization and coding
10-bit codes become 8-bit codes and a codec removes what is left in between, so shadows band.
A converter has to invent what the frame no longer holds. The design question is what to invent, and what to leave alone.
The frozen VAE in LumaFlux was trained on SDR. The tone decoder calibrates its output, but no method recovers detail that heavy clipping removed.
LumaFlux (arXiv:2604.02787)
Against the other generative converters
FR-HIDROVQA (lower is better) on 30 expert-track frames, with median seconds per frame above each bar.
FR-HIDROVQA is a full-reference distance on HIDRO-VQA features (Saini et al., WACV 2024).
Method
Steps
Params (M)
Trainable (M)
s/frame
Mem. (GB)
HDRTVNet++ (CNN, for scale)
1
0.59
0.59
0.857
2.31
LEDiff (5122)
50
1370.23
859.53
4.449
6.37
X2HDR (5122)
30
17020.02
149.42
4.244
33.77
LumaFlux (ours)
8
12484.77
71.32
5.282
27.06
One GH200, batch 1. LEDiff and X2HDR run at 5122; at 1080p scale, 35.19 and 33.57 s.
LumaFlux scores 0.431, 0.477 and 0.567 at 540p, 720p and 1080p, the lowest at each, in 2.64, 3.08 and 5.29 s.
That is 2.4 to 4.7× faster than X2HDR and 4.0 to 16.1× faster than LEDiff. HDRTVNet++ stays 173 to 201× faster.
All three generative models share one output contract: 99.5th luminance percentile anchored at 1,000 nit, BT.709 to BT.2020, clip, ST 2084.
LumaFlux loads no text encoder and takes 8 Euler steps.
LumaFlux (arXiv:2604.02787)
How LumaEval scores every method
Track
Pairs
What is scored
Checkpoint
Synthetic (primary)
100
20 held-out CHUG and LIVE-TMHDR contents, eight TMO variants, 78 video-variant groups
mixed-corpus
Expert
100
LIVE-TMHDR pairs whose SDR is an expert grade
mixed-corpus
Native
117
HDRTV1K test pairs
mixed-corpus
Legacy HDRTV1K
117
same pairs, five-metric suite, published baseline values
HDRTV1K-specific
Fixed-QP
468
intra-only HEVC on the SDR input at QP 27, 32, 37, 42
both
Temporal
5 clips
Bonfire and four CHUG clips, PU21 flicker energy
mixed-corpus
Text-to-HDR
6 clips
LTX-Video clips, 121 frames each, intrinsic metrics only
mixed-corpus
Outputs are PNG16 PQ BT.2020 at 1,000 nit, with identical inputs and metric code. CHUG: Saini et al., ICIP 2025.
Degradations
BT.2390 EETF, BT.2446 Method C and hard clip (each with a gamut map); Reinhard; BT.2446 Method A; gamma clip; OpenColorIO v2; YouTube LogC.
Metrics
PU21-PSNR (RGB and Y), PU21-SSIM, ΔEITP, HDR-LPIPS, HDR-VDP-3, and FR-HIDROVQA, a full-reference distance on HIDRO-VQA features (Saini et al., WACV 2024).
Baselines
Analytic BT.2446c and Reinhard; CNNs HDRTVNet++, FMNet, ITM-LUT, KUNet; PromptIR with a fixed Reinhard inverse; generative LEDiff (50 steps) and X2HDR (30 steps); a VAE + RQS control with no diffusion transport.
KUNet and PromptIR cover 86 of the 100 synthetic IDs, and LEDiff and X2HDR only five variants, so they stay off the primary table.
LumaFlux (arXiv:2604.02787)
Flow matching in four lines
The training recipe behind FLUX.1-dev, which LumaFlux keeps unchanged
xt = (1 − t) x0 + t x1
A straight path from a data sample x0 to a noise sample x1, indexed by t in [0, 1].
ut = x1 − x0
The target velocity along that path is constant.
L = E ‖ vθ(xt, t) − (x1 − x0) ‖2
Train the network vθ by least squares on the velocity.
dx / dt = vθ(x, t)
Sample by integrating from t = 1 back to t = 0 with a few Euler steps. No noise is injected on the way.
A text-to-image model starts at pure noise and needs many steps. LumaFlux starts much closer to the answer, which is the next slide.
LumaFlux (arXiv:2604.02787)
Rectified flow and the SDR-conditioned bridge
From Liu et al., Flow Straight and Fast. Rectification straightens crossing paths, so a few Euler steps track them closely.
Training interpolates zt = (1 − t) zHDR + t z1; the network predicts the straight-path velocity z1 − zHDR.
Sampling starts at the encoded SDR plus a small perturbation. There is no partial-noise start and no sampling-strength knob.
8 uniform Euler steps with fixed guidance: 5.28 s and 27.06 GB per 1080p frame on one GH200. 40 steps take 24.99 s, 4.7× the time, with no calibration gain.
Loss (5): (λ1, λ2, λ3) = (1, 0.5, 0.1). LFM regresses the bridge velocity, the L1 terms compare linear light at the 1,000-nit peak, Lsmooth regularizes adjacent knot slopes.
Decoded losses run every fourth optimizer step on the one-step estimate ẑ0 = zt − t v̂t; the sampler is never unrolled.
LumaFlux (arXiv:2604.02787)
The frozen backbone, and why plain LoRA is not enough
SigLIP-SO400M at 3842, giving 729 tokens of width 1152.
No text encoder is loaded. The text stream receives 8 learned null tokens plus the SigLIP tokens.
Plain LoRA
W = W(0) + A B, rank(AB) ≪ d
The same low-rank update for every token, every step and every block, whatever the frame contains.
Nothing maps the SDR-trained VAE output into HDR.
What the ablation says
Configuration (20k steps, 82 held-out frames)
PSNR
ΔEITP
(a) LoRA only, no tone decoder
19.29
256.84
(b) attention + MLP LoRA, with the RQS decoder
28.36
36.35
(g) full LumaFlux design
29.19
27.81
Higher PSNR and lower ΔEITP are better. The full ablation matrix is on the ablation slide.
Removing the tone decoder from the full design costs 9.69 dB. Conditioning the adaptation on what the frame measures adds 0.83 dB over plain LoRA and cuts color error by 8.53.
LumaFlux (arXiv:2604.02787)
The full architecture
Physical measurements and SigLIP tokens condition every frozen block; the frozen VAE and the RQS head produce PQ BT.2020.
z = z + γ ⊙ LN(z) + ζ # PCM
h = LN(z)
V_i = h_i (W_V + R_v,i) # PGA, per token
z = z + Attn(Q, K, V)
z = z + MLP(LN(z))
z = z + λ [W_p T_phys + W_c R(C T_perc)] # coupler
Rv,it,l = ( αpgat,l AvBv + βpgat,l I ) Gphys,i ( I + nspect,l Diag(gFFT) )
Every added path starts at zero (Bv, Wr, γ, ζ and the coupler's last layers), so step zero is exactly the frozen FLUX.
[αpga, βpga, αpcm, βpcm, nspec, λ] = Ψ(t, l): six gains per flow step and block, from sinusoidal flow-time features and learned block embeddings, 41,094 parameters.
Y = m2020T xlin with m2020 = [0.2627, 0.6780, 0.0593]; Tphys = Conv3×3([Y, log(1 + |∇Y|), sat]); g = MLPg([μY, σY, p95, p99]) in ℝ16; r has Kr = 8 radial bands.
LumaFlux (arXiv:2604.02787)
Where the highlights go
Headroom maps of one LumaEval frame. Gray is at or below 100 nit; color marks pixels above the SDR ceiling.
Method
Share above 100 nit
IoU with reference highlights
Error inside highlights (stops)
Reference
32.3%
BT.2446c
19.1%
Reinhard
29.7%
0.563
HDRTVNet++
47.9%
0.538
0.81
LumaFlux
37.5%
0.579
0.87
Means over 100 synthetic-track frames; percentages under the maps are for that frame. Higher IoU is better.
The analytic curve under-expands, HDRTVNet++ pushes too much of the frame bright, and LumaFlux puts highlights where the reference has them.
LumaFlux (arXiv:2604.02787)
What CFG does to flow models
Every image here is standard CFG on a flow model. These are typical results; none were cherry-picked.
Flux-dev with CFG, ω = 3 (paper Figure 6)
Plastic, cartoon look
Oversaturated
Flat, clipart look
Blown highlights
Washed out
SD3 with CFG, ω = 3 to 3.5
Oversaturated
Text garbled
Misspelled
Color blow-out
Color distortion
Four failure modes. Oversaturation (colors pushed past their natural range), cartoonification (photoreal turns plastic), text corruption (letters distort and merge), and detail blow-out (fine texture becomes flat patches).
Not rare. CFG artifacts show up on every prompt across Flux, SD3 and SD3.5. The cause is structural: CFG extrapolation is incompatible with the deterministic ODE that every modern flow model uses.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
What good guidance for a rectified flow should do
D1. Stay on the manifold. The guided trajectory stays in a bounded neighbourhood of the learned transport manifold Mt. No off-manifold drift, no error accumulation.
D2. Strong early, gentle late. Early steps set the global structure (layout, color palette), so guidance should be strong. Late steps resolve fine detail (text, texture), so guidance should vanish to avoid corrupting it.
D3. Provable guarantees. Formal bounds on manifold distance and on how far the output distribution drifts, on top of looking better. No prior guidance method has them.
D4. Drop-in replacement. Works with any pretrained flow model (Flux, SD3, SD3.5, Lumina) without retraining. Swap the sampling algorithm and nothing else.
Do existing methods satisfy these?
Method
D1
D2
D3
D4
CFG
no
no
no
yes
CFG-Zero*
no
no
no
yes
APG
partly
no
no
yes
Rectified-CFG++ (ours)
yes
yes
yes
yes
Our idea: predict a midpoint on the manifold, evaluate guidance there, and add it to the conditional velocity as a bounded correction, in place of extrapolating from the current point.
Three ideas, three guarantees. A predictor half-step gives on-manifold evaluation, an adaptive α(t) gives vanishing guidance, and a bounded correction gives interpolation in place of extrapolation.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
The ideal flow on the manifold (1 of 3)
What you are looking at. Three manifolds: Mt at the current time (the blue and gold band), Mt−1 at the next step, and M0, the data (the orange curve at the bottom). The red dots are the ideal positions at each timestep.
The ideal trajectory. The sample flows along the manifold surface from Mt to Mt−1 to M0. At each step the velocity lies in the tangent space of the manifold and never leaves it.
Three flows. Blue is the Flux conditional velocity vc, red is CFG (extrapolated), green is Rectified-CFG++ (ours). Green and blue stay near the manifold while red diverges.
The question. How do we guide the generation for better text adherence while staying on the manifold? First, what goes wrong.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
CFG extrapolates off the manifold (2 of 3)
What CFG does. Starting from the black dot on Mt, CFG computes the guided velocity v̂ω at the current point. With ω > 1 this velocity extrapolates past the conditional velocity vc.
v̂ω = vu + ω(vc − vu), ω > 1 overshoots
The tangent space violation.v̂ω does not belong to the tangent space TxtMt. The Euler step lands the sample off Mt−1, and in a deterministic ODE there is no way back.
The compounding effect. At the next step CFG evaluates guidance at the off-manifold point, which makes the direction worse again. Errors accumulate at every step. In the final image: oversaturation, color blow-out, structural distortion.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Our predictor-corrector stays on the manifold (3 of 3)
Predict a midpoint with vc only. Take a half-step with the conditional velocity: x̃mid = xt + (Δt/2) vct. Since vc lies in the tangent space, x̃mid lands on or near Mt−Δt/2.
Evaluate guidance at the midpoint. Compute vcmid and vumid at x̃mid. The guidance direction Δvmid = vcmid − vumid is evaluated on the manifold, so it reflects the geometry of the target manifold Mt−Δt.
Apply a bounded interpolative correction.
v̂ = vct + α(t)(vcmid − vumid)
α(t) = λmax(1−t)γ. The base velocity is always vc, which is on the manifold. The correction is additive and bounded, and α(t) goes to 0 near the data, so fine detail survives.
The trajectory stays in a bounded tubular neighbourhood of Mt.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Four guidance rules for the same model
Only ours avoids extrapolation altogether
Standard CFG (Ho and Salimans, 2022)
v̂ = (1 − ω) vu(xt) + ω vc(xt)
ω > 1 extrapolates past vc. Designed for SDEs with renoising. On an ODE: off-manifold drift, oversaturation, error accumulation. Constant guidance at every timestep.
CFG-Zero* (Fan et al., 2025)
v̂ = (1 − ω) st*vu(xt) + ω vc(xt)
Adds an optimal rescaling factor st* to align the unconditional and conditional norms. Less initial drift, but it still extrapolates (ω > 1) and has no late-step adaptation, so fine detail is still corrupted.
APG (Sadat et al., 2024)
v̂ = vc(xt) + Δvt(η, r, β)
Projects the guidance direction onto the perpendicular subspace to shrink the parallel component. Partial mitigation, but guidance is still evaluated at the current point (possibly off the manifold). No decay schedule, no formal bounds.
Rectified-CFG++ (ours)
v̂ = vc(xt) + α(t) (vcmid − vumid)
Base is always vc (on the manifold). Guidance at the predicted midpoint x̃mid (on the manifold). α(t) = λ(1−t)γ goes to 0 near the data. Bounded, interpolative, provable.
Property
CFG
CFG-Zero*
APG
Ours
Guidance type
extrapolation
extrapolation
projection
interpolation
Evaluated at
current xt
current xt
current xt
predicted x̃mid
Temporal schedule
none (constant)
none
none
α(t) decays to 0
Formal guarantees
no
no
no
yes (3 results)
NFE per step
2
2
2
3
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
A 2D toy example
CFG drifts off the support, ours stays on the manifold
200 trajectories on a 2D mixed Gaussian. Top: CFG (blue) drifts off the support. Bottom: ours (green) moves smoothly along the manifold.
CFG (top). Trajectories overshoot at first and leave the learned transport manifold. Sharp late corrections pull the samples back, but the damage is done. The final distribution is noisy and off-center.
Rectified-CFG++ (bottom). Smooth convergence throughout. The predictor anchors each step on the flow field and the corrector applies bounded guidance. Samples arrive at the target tightly concentrated, with no drift and no sharp corrections.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Same prompt, same seed, three guidance strategies on Flux
Flux-dev, same prompt, same seed. Left: no guidance (baseline). Center: CFG, ω = 3.5. Right: Rectified-CFG++, λ = 0.5. Ours has sharper detail, faithful color and no rainbow artifacts.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Rectified-CFG++ on random prompts
All images from Rectified-CFG++ on Flux-dev. Random prompts from MS-COCO and Pick-a-Pic, no cherry-picking.
Photorealistic scenes. Landscapes, portraits, food, architecture. Natural lighting, correct shadows, no oversaturation. Staying on the manifold preserves the photoreal distribution Flux was trained on.
Artistic and stylized. Oil paintings, digital art, fantasy scenes. Style transfer works because α(t) allows strong guidance early (global style) while preserving fine brush strokes late.
Text and composition. Signage, book titles, multi-object scenes, the hardest category for any guidance method. α(t) going to 0 at late steps preserves pixel-precise text and spatial relationships.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Four guidance methods on SD3.5, same prompt and seed
CFG, ω = 3.5
CFG-Zero*
APG
Ours, λ = 6
CFG
CFG
Ours
Ours
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
The same on SD3 and Lumina-Next
SD3, CFG
SD3, ours
SD3, CFG
SD3, ours
Lumina, CFG
Lumina, ours
Lumina, CFG
Lumina, ours
A drop-in replacement across rectified flow models. No retraining, no architecture changes.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Text rendering, a neon street sign
CFG
Rectified-CFG++
Prompt: "A neon street sign that says 'CyberCore Cafe', glowing in magenta and blue."
CFG garbles the neon letters into "Cyberre Cidie Cafe". The glow halos and letter boundaries bleed together. Ours renders each letter distinctly with clean separation between the neon glows.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Text rendering, headlines and inscriptions
CFG
Ours
CFG
Ours
Prompt: "A crow detective reading a paper titled 'Feathered Conspiracies', headline in bold noir font."
Prompt: "A magical sword embedded in stone, with the name 'SOLARFANG' etched along its blade in glowing runes."
CFG turns "Feathered Conspiracies" into "Feathhrad Conspiracies" and SOLARFANG into garbled runes. Ours renders both headlines correctly with proper typography, even on surfaces like aged paper and glowing stone.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Text rendering, a slide and a scroll
CFG
Ours
CFG
Ours
Prompt: "A fox giving a TED talk, slide behind reads 'Entropy and the Soul'."
Prompt: "A burning scroll with the title 'Soulbound by Nightfall' in ornate calligraphy."
CFG turns "Entropy and the Soul" into "Entφy amd the Soul", and the scroll title is barely readable. Ours gives clean title text on both the presentation screen and the burning scroll; zero guidance at the final steps preserves the fine detail.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Text rendering, carved and printed
CFG
Ours
CFG
Ours
Prompt: "A desert gate carved in sandstone reading 'CITY OF WHISPERS'."
Prompt: "A Japanese fish stamp print with 'WAVES OF JUSTICE' in red ink."
The pattern. Across neon, carved stone, ink stamps, calligraphy and digital displays, Rectified-CFG++ keeps the text legible. The adaptive schedule α(t) going to 0 is the mechanism: no guidance perturbation while fine detail is resolved.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Text legibility on SD3.5, the full gallery
Each pair: CFG on the left, Rectified-CFG++ on the right
CFG
Ours
CFG
Ours
CFG
Ours
CFG
Ours
CFG
Ours
CFG
Ours
CFG
Ours
CFG
Ours
Legible text on every prompt. The adaptive schedule α(t) is what does it.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)
Sweeping the guidance scale
CFG collapses at high guidance, ours stays flat
x-axis: guidance scale (λ for ours, ω for CFG). y-axis: FID ↓, CLIP ↑, ImageReward ↑, Aesthetic ↑. Blue is CFG, green is ours.
CFG (blue curves). FID explodes from 77 to 148 at ω = 10. ImageReward collapses from 1.0 to −1.5. CLIP drops 30%, Aesthetic drops 35%. Catastrophic failure at high guidance.
Rectified-CFG++ (green curves). FID stays at 75, flat. ImageReward goes from 1.0 to 0.95. CLIP and Aesthetic stay stable. Graceful degradation at every scale.
Why this matters. CFG needs careful per-model tuning of ω to avoid collapse. Rectified-CFG++ holds across a 50× range of λ with no hyperparameter sensitivity. One setting works everywhere.
PhD defenseRectified-CFG++ (Saini et al., NeurIPS 2025)