September 22, 2025 • 7 mins
Get ready for your fastest video edit yet. Today, Zane and Pippa break down Lucy Edit Dev from Decart AI—an open-weight, instruction-following video editor now on Hugging Face. Unlike text-to-video generators, Lucy Edit Dev lets you edit existing footage with natural language prompts: swap a kimono, replace a background, or turn a barista into an astronaut, all without painstaking masking. Powered by a VAE and Diffusion Transformer backbone in the tradition of Wan 2.2, Lucy Edit Dev keeps shot composition, motion, and timing stable while delivering precise instruction-driven edits. It is dev-focused, open for research and R and D, but under a non-commercial license, making it perfect for prototyping, benchmarking, or personal creative exploration. Find out why creators, marketers, designers, and researchers are excited about this new “do what I said” tool for efficient A B testing and visual experimentation. We cover setup needs, real-world uses, limitations like GPU compute, edge cases and license restrictions, plus how Lucy Edit Dev stacks up against Runway, Pika, Luma, and Adobe’s video AI tools. Whether you are tired of tedious rotoscoping or want to quickly generate multiple style options for your YouTube video, discover if Lucy Edit Dev unlocks a new layer of creative control and speed. Listen in for takeaways, practical tips, and the tagline memes you did not know you needed.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
Happy Monday, and welcome back to the Blue Lightning AI
Daily Podcast for Menday, September twenty second, twenty twenty five.
I'm Zane and yes, this episode was fully built with AI.
If you hear a weird glitch, we left it in
on purpose. Call it vibe.
Speaker 3 (00:13):
Enhancement, Big Vibe.
Speaker 2 (00:15):
I'm Pippa, caffeinated and curious, and today we're talking Lucy
Edit dev from dicart Ai. It's an open waight instruction
following video editor. You can grab on Hugging Face. You
type ad a patterned kimono. It does the wardrobe change
without you masking, frame by frame.
Speaker 1 (00:31):
That's the pitch, right. It's not a text to video generator.
It edits footage you already shot. According to the model
card on Hugging Face. The sweet spot is edits like
wardrobe swaps, character replacements, object insertions or removals, and full
background changes while preserving motion and composition source on That
is the dcarte dash Ai dash lad Licensing youton hef
edit dev model.
Speaker 2 (00:51):
Card, which matters because roto is pain. If you're a
solo YouTuber or a brand team punching out ab variants,
this is ours saved make the jacket read, drop her
into a Neon city, swap the barista for an astronaut.
You brief it like a human editor.
Speaker 1 (01:08):
Context check what's actually new here. It's the instruction following
part tune for edits not prompts for generation. The model
emphasizes verbs ad replace transform, so your intent is legible,
and it tries hard to keep your blocking, timing and
camera moves intact.
Speaker 2 (01:23):
Also, it's open weights with a non commercial license. That
combo's kind of the headline for devs and researchers. You
can run it locally, test it in your pipeline, but
you can't ship it in a paid product. Research First
Vibes hugging Face has the weights.
Speaker 1 (01:37):
We saw it there under the hood. Descartes says it's
a VIAE plus a diffusion transformer backbone tied to the
WAND two point two five B lineage. If you've tracked
wand two point two's open research on temporal coherence that
explains why motion looks stable here. They reference the wand
video WAND two point two GitHub repo in their materials.
Speaker 3 (01:56):
Okay, who's this really for today?
Speaker 2 (01:58):
I'm seeing fashion and creators swap looks without reshoots, product marketers,
quick localized variants, change label color, add a holiday prop,
music video folks, surreal styling and background flips that still
respect choreography researchers and indie tool makers benchmarking, instruction, fidelity
and drift.
Speaker 1 (02:15):
And because it's dev oriented, expect some set up Python
diffusers GPU, the model cards examples, show short clips, social
friendly durations and inference on kudah. If you've run other
diffusion video tools, this lands in that same complexity zone.
Speaker 2 (02:29):
Wait, so is this a big game changer or a
nice to have? I think for editing existing footage with
natural language. It's a meaningful step for full time creators
who hate keyframes. It's a workflow unlock for casual hobbyists.
Kind of spicy, but you'll want a UI later.
Speaker 1 (02:45):
Yeah, it won't replace Premiere or Resolve, but it might
replace that open after effects and cry step when you
just need one visual change across sixty frames, direction over dials.
Speaker 2 (02:54):
Competitive field check, Runway, Pica, Luma, Adobe, All Play, and
Video AI. But most tools lead with generation or closed
hosted editors. Lucy Edit dev stands out because it's open
weights and specifically instruction following for edits. It's not trying
to be your whole editor, it's trying to be the
do what I said layer, and it's early.
Speaker 1 (03:12):
I haven't seen formal quantitative benchmarks on the model card,
no big psn R charts or whatever, so consider this
a hands on and see release. But the examples do
show consistent motion with wardrobe and background swaps. Again Source
Hugging Face model card for Descartes dash Ai Dash.
Speaker 2 (03:27):
Lucy, Edit dev Pop quiz Zane mean this update.
Speaker 1 (03:31):
Stylus called in sick Lucy did the fittings you We
fix it.
Speaker 2 (03:36):
In post, but the post is a genie one wish
per shot.
Speaker 1 (03:39):
Let's talk workflow impact. A solo YouTuber shooting b roll
for a product review could swap packaging color for regional
variants in minutes, remove a stray logo on a shirt
without masking every frame, test backgrounds to match a brand
palette before moving to heavy post.
Speaker 2 (03:54):
Podcasters doing video imagine changing your virtual set without green screen.
Make it a cozy brick wall done Tiktoker's style flips
mid dance, new jacket, new hat, same moves. Designers quick
motion tests for clients, drop the model into the campaign
world before a location.
Speaker 1 (04:10):
Shoot and for filmmakers, I'd call this a previs superpower.
You can pitch a look nineteen fifties noir suit rainy
Alley without tying up VFX. If the client bites, then
you shoot or do the full.
Speaker 2 (04:21):
Post limitations time, non commercial license, huge guardrail. Great for
R and D, not for selling edits. Also, like any
diffusion video edit, edge cases will artifact thin straps, hand
object interactions, fast motion with motion blur, expect hiccups. Brand
safety and likeness rights still apply and compute.
Speaker 1 (04:40):
This is a modern video diffusion stack plan on a
decent GPU and patience. It's not click once on your phone,
at least not.
Speaker 2 (04:46):
Yet availability live now on hugging face as Descartes dash
Ai dot Ai at a stock Debije systems.
Speaker 3 (04:53):
The card to y sundongenbus on this check to much fifth.
Speaker 2 (04:57):
Like the anglis t trend wi much, we're moving from
prompt to generate to instructions to edit. It's less make
a new video and more respect the shot I already crafted.
That's huge for directors who care about lensing, blocking, lighting,
all the stuff you dialed in on set.
Speaker 1 (05:13):
Quick comparisons versus runway or Pika. Those can do video
to video and style transfers, but Lucy Edit Dev's pitch
is tight instruction fidelity on existing footage and being open
weight for local eval versus Adobe's genfill for video, Adobe's
Integrated closed and UI polished. This is flexible and hackable, but.
Speaker 3 (05:31):
Dev facing price and value.
Speaker 2 (05:34):
The weights are free to download for research, but your
cost is GPU time and the non commercial restriction. If
you need to ship, look at pro hosted options our
weight for a commercial friendly license risks.
Speaker 1 (05:44):
Watermarks aren't the issue here. Licensing and creative control are
also disclosure. If you're swapping wardrobe or backgrounds, note your
brand or platform policies and keep an eye on identity
drift in character swaps when faces turn or acclude.
Speaker 2 (05:58):
Okay, real talk. How would we use it? I'd shoot
a quick dance loop, then try five looks streetwear, vintage, neon, cyberpunk,
same motion, different vibe, instant mood board video.
Speaker 1 (06:07):
I'd use it for product tabletop swap label variants and
background textures. Send the client three options before doing a
full res pass in traditional post.
Speaker 2 (06:16):
If you're listening and want to test, grab a clean
medium shot with stable lighting. Limited occlusions, write clear instructions
with verbs and nouns. Replace denim jacket with glossy black
leather biker jacket, and keep the first test short.
Speaker 3 (06:29):
If it works, go bigger.
Speaker 1 (06:31):
One more note for devs. The architecture tie to one
two point two suggest good temporal handling, but your pre
processing matters consistent frame rate, sensible resolution. The example code
on the model card uses KUDA and short clips. Start there,
and please.
Speaker 2 (06:45):
Don't ask about token caps. Wrong party, bestie. This isn't
an LM. It's a diffusion video editor, different brain.
Speaker 1 (06:52):
Big Picture. Lucy edit dev is a signal editing what
you shot with natural language. Direction is landing in practical workflows,
open for research, a pro tier for production, classic funnel.
Speaker 2 (07:04):
All right, that's our take today's episode, AI made Tiny
glitches and all covered. Lucy edit dev on hugging face
and the one two point two lineage via GitHub. If
you missed our recent chats on runways, control updates or
Luma's latest peep the Blue Lightning blog.
Speaker 1 (07:20):
Thanks for hanging out with us on the Blue Lightning
AI Daily podcast. For news updates and tutorials on your
favorite AI tools, head to blue lightningtv dot com catch
you next time.
Speaker 2 (07:29):
Be good, make cool stuff and fix it in post,
just with fewer tears.
Speaker 3 (07:33):
Bye.
Happy Monday, and welcome back to the Blue Lightning AI
Daily Podcast for Menday, September twenty second, twenty twenty five.
I'm Zane and yes, this episode was fully built with AI.
If you hear a weird glitch, we left it in
on purpose. Call it vibe.
Speaker 3 (00:13):
Enhancement, Big Vibe.
Speaker 2 (00:15):
I'm Pippa, caffeinated and curious, and today we're talking Lucy
Edit dev from dicart Ai. It's an open waight instruction
following video editor. You can grab on Hugging Face. You
type ad a patterned kimono. It does the wardrobe change
without you masking, frame by frame.
Speaker 1 (00:31):
That's the pitch, right. It's not a text to video generator.
It edits footage you already shot. According to the model
card on Hugging Face. The sweet spot is edits like
wardrobe swaps, character replacements, object insertions or removals, and full
background changes while preserving motion and composition source on That
is the dcarte dash Ai dash lad Licensing youton hef
edit dev model.
Speaker 2 (00:51):
Card, which matters because roto is pain. If you're a
solo YouTuber or a brand team punching out ab variants,
this is ours saved make the jacket read, drop her
into a Neon city, swap the barista for an astronaut.
You brief it like a human editor.
Speaker 1 (01:08):
Context check what's actually new here. It's the instruction following
part tune for edits not prompts for generation. The model
emphasizes verbs ad replace transform, so your intent is legible,
and it tries hard to keep your blocking, timing and
camera moves intact.
Speaker 2 (01:23):
Also, it's open weights with a non commercial license. That
combo's kind of the headline for devs and researchers. You
can run it locally, test it in your pipeline, but
you can't ship it in a paid product. Research First
Vibes hugging Face has the weights.
Speaker 1 (01:37):
We saw it there under the hood. Descartes says it's
a VIAE plus a diffusion transformer backbone tied to the
WAND two point two five B lineage. If you've tracked
wand two point two's open research on temporal coherence that
explains why motion looks stable here. They reference the wand
video WAND two point two GitHub repo in their materials.
Speaker 3 (01:56):
Okay, who's this really for today?
Speaker 2 (01:58):
I'm seeing fashion and creators swap looks without reshoots, product marketers,
quick localized variants, change label color, add a holiday prop,
music video folks, surreal styling and background flips that still
respect choreography researchers and indie tool makers benchmarking, instruction, fidelity
and drift.
Speaker 1 (02:15):
And because it's dev oriented, expect some set up Python
diffusers GPU, the model cards examples, show short clips, social
friendly durations and inference on kudah. If you've run other
diffusion video tools, this lands in that same complexity zone.
Speaker 2 (02:29):
Wait, so is this a big game changer or a
nice to have? I think for editing existing footage with
natural language. It's a meaningful step for full time creators
who hate keyframes. It's a workflow unlock for casual hobbyists.
Kind of spicy, but you'll want a UI later.
Speaker 1 (02:45):
Yeah, it won't replace Premiere or Resolve, but it might
replace that open after effects and cry step when you
just need one visual change across sixty frames, direction over dials.
Speaker 2 (02:54):
Competitive field check, Runway, Pica, Luma, Adobe, All Play, and
Video AI. But most tools lead with generation or closed
hosted editors. Lucy Edit dev stands out because it's open
weights and specifically instruction following for edits. It's not trying
to be your whole editor, it's trying to be the
do what I said layer, and it's early.
Speaker 1 (03:12):
I haven't seen formal quantitative benchmarks on the model card,
no big psn R charts or whatever, so consider this
a hands on and see release. But the examples do
show consistent motion with wardrobe and background swaps. Again Source
Hugging Face model card for Descartes dash Ai Dash.
Speaker 2 (03:27):
Lucy, Edit dev Pop quiz Zane mean this update.
Speaker 1 (03:31):
Stylus called in sick Lucy did the fittings you We
fix it.
Speaker 2 (03:36):
In post, but the post is a genie one wish
per shot.
Speaker 1 (03:39):
Let's talk workflow impact. A solo YouTuber shooting b roll
for a product review could swap packaging color for regional
variants in minutes, remove a stray logo on a shirt
without masking every frame, test backgrounds to match a brand
palette before moving to heavy post.
Speaker 2 (03:54):
Podcasters doing video imagine changing your virtual set without green screen.
Make it a cozy brick wall done Tiktoker's style flips
mid dance, new jacket, new hat, same moves. Designers quick
motion tests for clients, drop the model into the campaign
world before a location.
Speaker 1 (04:10):
Shoot and for filmmakers, I'd call this a previs superpower.
You can pitch a look nineteen fifties noir suit rainy
Alley without tying up VFX. If the client bites, then
you shoot or do the full.
Speaker 2 (04:21):
Post limitations time, non commercial license, huge guardrail. Great for
R and D, not for selling edits. Also, like any
diffusion video edit, edge cases will artifact thin straps, hand
object interactions, fast motion with motion blur, expect hiccups. Brand
safety and likeness rights still apply and compute.
Speaker 1 (04:40):
This is a modern video diffusion stack plan on a
decent GPU and patience. It's not click once on your phone,
at least not.
Speaker 2 (04:46):
Yet availability live now on hugging face as Descartes dash
Ai dot Ai at a stock Debije systems.
Speaker 3 (04:53):
The card to y sundongenbus on this check to much fifth.
Speaker 2 (04:57):
Like the anglis t trend wi much, we're moving from
prompt to generate to instructions to edit. It's less make
a new video and more respect the shot I already crafted.
That's huge for directors who care about lensing, blocking, lighting,
all the stuff you dialed in on set.
Speaker 1 (05:13):
Quick comparisons versus runway or Pika. Those can do video
to video and style transfers, but Lucy Edit Dev's pitch
is tight instruction fidelity on existing footage and being open
weight for local eval versus Adobe's genfill for video, Adobe's
Integrated closed and UI polished. This is flexible and hackable, but.
Speaker 3 (05:31):
Dev facing price and value.
Speaker 2 (05:34):
The weights are free to download for research, but your
cost is GPU time and the non commercial restriction. If
you need to ship, look at pro hosted options our
weight for a commercial friendly license risks.
Speaker 1 (05:44):
Watermarks aren't the issue here. Licensing and creative control are
also disclosure. If you're swapping wardrobe or backgrounds, note your
brand or platform policies and keep an eye on identity
drift in character swaps when faces turn or acclude.
Speaker 2 (05:58):
Okay, real talk. How would we use it? I'd shoot
a quick dance loop, then try five looks streetwear, vintage, neon, cyberpunk,
same motion, different vibe, instant mood board video.
Speaker 1 (06:07):
I'd use it for product tabletop swap label variants and
background textures. Send the client three options before doing a
full res pass in traditional post.
Speaker 2 (06:16):
If you're listening and want to test, grab a clean
medium shot with stable lighting. Limited occlusions, write clear instructions
with verbs and nouns. Replace denim jacket with glossy black
leather biker jacket, and keep the first test short.
Speaker 3 (06:29):
If it works, go bigger.
Speaker 1 (06:31):
One more note for devs. The architecture tie to one
two point two suggest good temporal handling, but your pre
processing matters consistent frame rate, sensible resolution. The example code
on the model card uses KUDA and short clips. Start there,
and please.
Speaker 2 (06:45):
Don't ask about token caps. Wrong party, bestie. This isn't
an LM. It's a diffusion video editor, different brain.
Speaker 1 (06:52):
Big Picture. Lucy edit dev is a signal editing what
you shot with natural language. Direction is landing in practical workflows,
open for research, a pro tier for production, classic funnel.
Speaker 2 (07:04):
All right, that's our take today's episode, AI made Tiny
glitches and all covered. Lucy edit dev on hugging face
and the one two point two lineage via GitHub. If
you missed our recent chats on runways, control updates or
Luma's latest peep the Blue Lightning blog.
Speaker 1 (07:20):
Thanks for hanging out with us on the Blue Lightning
AI Daily podcast. For news updates and tutorials on your
favorite AI tools, head to blue lightningtv dot com catch
you next time.
Speaker 2 (07:29):
Be good, make cool stuff and fix it in post,
just with fewer tears.
Speaker 3 (07:33):
Bye.
Popular Podcasts
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
The Joe Rogan Experience
The official podcast of comedian Joe Rogan.